Category: 科学易学

  • 那个决定门槛高低的系数不是猜的:它可由读者各自的边际直接算出,而它并不只是 p_e 的函数

    Evidence level: C

    门槛公式 n ≈ C × (39.2/(1−p_e))² 里唯一被当成「数值读数」的 C,其实有闭式解;但把它化简成 p_e 的单变量函数会引入两到四倍半的误差,因为真 C 由两个读者各自的整条边际共同决定。

    正文

    一、问题的形状:一个被当成读数的数

    先接上一篇的尾巴。#1409 把判读一致性天花板的样本量门槛压成了一条合式:所需例数约等于 C 乘以(39.2 除以 1 减 p_e)的平方。其中放大因子 1/(1−p_e) 是精确代数,没有一个符号是估的;但 C 只是五个 m̄ 取值上的数值读数——0.60 与 0.75 附近约 0.42,0.90 约 0.25,0.95 约 0.18。这篇要回答的是:C 到底是什么的函数。

    二、C 是两读者边际的直接函数

    第一步先写下 C 的定义。TVD 是两读者各档比例的绝对差之和的一半,在大样本下它的抽样波动可以线性化:每一档的系数取该档差值的符号的一半,于是 C 等于这个符号向量在「两读者比例协方差矩阵之和」上的二次型。符号向量与协方差矩阵都只依赖两个读者的比例本身,所以 C 从来不是自由参数——它是两个边际的直接函数。真正的问题因此不是「C 能不能算」,而是「算它需要几个数」。

    三、三个单参数简化式全部失败

    这里有个陷阱。逻辑上最自然的做法是假设 C 只由 p_e 决定,毕竟门槛公式里 p_e 是唯一出现在分母上的分布量;文献也反复强调 κ 的标准误依赖边际,但往往把「边际」当成一个笼统的难度旋钮。我们把三类单参数简化式都试了:只用读者一的主档、用两读者主档的平均、只用 1 减 p_e。在 K 从 2 到 6、p_e 从 0.36 到 0.90、真实 TVD 从 0 到 0.20、差落在近端或远端或两档分摊的三格局共 360 个格点上,这些简化式落在真值正负一成以内的只有 12% 到 19%;中位绝对偏差三成到五成,最坏处偏差四点五倍。换句话说,用 p_e 单变量去代 C,误差不是小数点后的事。

    四、零差档的折叠正态修正

    第二个陷阱更隐蔽。朴素线性化有个失效点:当两读者在某一档上的真实比例完全相等时,该档的符号为零,于是这一档的抽样波动被整个抹掉。但「真实相等」并不意味着「观测相等」——这一档的观测差在零附近来回跳,绝对值并不为零,它贡献的是折叠正态的方差:零点二五乘以(1 减 2/π)再乘该档方差。我们按这个修正重写了 C,得到一条只依赖两读者全部边际的闭式。在 54 个格点、每格五万一千二百例、一万次重抽样的核对中,闭式与蒙特卡洛的相对误差中位不到百分之一,最大 2.5%,全部落在 5% 以内。

    五、C 的真身:边际形状与差落点的联合函数

    修正之后,C 的结构就清楚了。它既不是 p_e 的函数,也不是 m̄ 的函数,而是「边际形状 + 差落点」的联合函数。固定 p_e 与真实 TVD 完全不变、只改差落在哪几档,C 的跨度在 K=3 时可达 1.32 倍,K=6 时 1.19 倍。p_e=0.65、K=5、TVD=0.10 的例子更直观:差全落在一条主档与一条末档之间时 C 是 0.2346,门槛约两千九百例;同样的边际、同样的 TVD,但差由两条档各分摊一半时 C 升到 0.2823,门槛三千五百例。两个真实分歧程度完全一样的研究,仅因为「分歧集中在一条线上」还是「摊在两条线上」,所需样本量差两成。

    那为什么 #1409 那五个 m̄ 读数看起来单调下降、很规整?因为那五个点全是同一种构型——差集中在主档与次档之间,而边际也随之从平缓变尖锐。单调下降是「构型固定、只扫边际」的产物,不是 C 的内在律。一旦构型换掉,同样的 p_e 下 C 可以反向。这解释了低一致区那处反常:0.60 与 0.75 平在同一水平,不是测量噪声,而是两个机制在此交点处互相抵消——边际越平缓,逐档比例越接近,符号向量的非零分量越多,方差上界抬高;而同时各档差值变小,真实波动收窄。

    六、这对实践是减法

    这条结果对实践的含义是减法而非加法。第一,门槛可以在招募前算出来,但前提是报全两个读者各自的期望边际分布,而不是只报一个 p_e 或一个 κ 目标值——这一点与方法学文献的要求一致:按目标精度反解样本量时必须指定完整边际,只给系数无法定 n。第二,任何「用小样本试点估一个 C 再外推」的路线都不成立,因为 C 随构型的敏感度与 n 无关,试点给不出构型信息。第三,若实践上只能给一个数,那这个数应当按最不利构型取上界——在本轮网格里,两档分摊构型普遍给出更高的 C,用它做保守估计不会低估门槛。

    证据与局限

    已验证(纯代数与蒙特卡洛,可复现):TVD 的线性化方差只依赖两读者边际,符号向量与协方差矩阵可显式写出;零差档必须按折叠正态补 0.25×(1−2/π)×方差项,否则朴素线性化系统性低估,零差档数从零增到三时低估幅度中位从 0% 增至 12.4%,最坏 17%;修正后的闭式在 54 格点、五万例规模下与仿真相对误差中位 0.9%、最大 2.5%。

    部分验证:C 对「差落点」的敏感度(同 p_e、同 TVD 下跨 1.14 至 1.32 倍)只在三档构型(远端单一、近端单一、两档分摊)上扫过,未穷举全部支撑集;「两档分摊构型给出更高 C」这一规律在 K=3 到 6、p_e=0.40 到 0.80 的九个格点上一致,但未证明对所有构型成立,只能作保守上界的经验依据而非定理。

    待验证:真实判读研究里「差落点」的分布从未被系统报告过——已发表研究普遍只给 κ、AC1 与 95% 区间,很少给两读者各自的完整边际,因此无法反推构型;这既是本条结论的适用边界,也是下一项工作的入口。此外 C 的闭式在 K 大于 6 或存在结构性零格(某档两读者都从未使用)时是否仍成立,本轮未测。

    L1 网络状态:L1 边表与对应网络文件中无「判读一致性度量的方差结构」这条边。最接近的是 E-060 标签前提(状态 conflict:PMID 30383398 报中医医师间诊断一致性在开放人群中为 low、作者称「令人担忧」;KB-317/334 报 AC2 已在真实临床场景跑通、并指出「患病率 ≠ 评定者所见患病率」的实操坑;KB-333 报 κ 型系数对序数/名义的敏感性方向与 Gwet 一系相反),该边只登记「标签信度够不够用」,未登记「信度指标的方差由什么决定」。故本文封顶 C,且不主张任何已验证机制。该主题尚未进入 L1 对应网络,缺口已登记。

    相关阅读

    证据等级: C
    参考文献: PMID 11512137, PMID 22560852, PMID 36123642, PMID 38283651, PMID 23627889, PMID 25168681, PMID 15733050


    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 coefficient that determines the threshold is not guessed: it can be calculated directly from each reader’s marginal utility, and it is not merely a function of p_e.

    Evidence level: C

    In the threshold formula $n \approx C \times (39.2/(1-p_e))^2$, the only term treated as a “numerical reading” is $C$. However, $C$ actually possesses a closed-form solution; simplifying it into a single-variable function of $p_e$ introduces an error ranging from two to four and a half times, because the true $C$ is jointly determined by the entire margin of each of the two readers.

    Main Text

    I. The Shape of the Problem: A Number Treated as a Reading

    To continue from the previous section. Reference #1409 compressed the sample size threshold for judging consistency into a tractable form: the required number of cases is approximately equal to $C$ multiplied by the square of (39.2 divided by 1 minus $p_e$). Among these, the amplification factor $1/(1-p_e)$ is exact algebra; no symbol here is estimated. However, $C$ itself is merely a numerical reading taken from five values of $\bar{m}$—approximately 0.42 near 0.60 and 0.75, approximately 0.25 near 0.90, and approximately 0.18 near 0.95. This paper addresses the question: what is $C$ actually a function of?

    II. $C$ is the Direct Function of Two Readers’ Margins

    The first step is to write down the definition of $C$. TVD (Traditional Consensus) is half the sum of absolute differences between the proportions of each category for the two readers. Under large sample conditions, its sampling fluctuation can be linearized: the coefficient for each category takes half the sign of the difference in that category. Thus, $C$ equals the quadratic form of this sign vector on “the sum of the two readers’ proportion covariance matrices.” Both the sign vector and the covariance matrix depend solely on the proportions of the two readers themselves. Therefore, $C$ is never a free parameter—it is a direct function of the two margins. The real problem is therefore not “Can $C$ be calculated?” but “How many numbers are needed to calculate it?”

    III. All Three Single-Parameter Simplifications Fail

    Here lies a trap. Logically, the most natural assumption is that $C$ depends only on $p_e$, since in the threshold formula $p_e$ is the only distributional quantity appearing in the denominator; literature also repeatedly emphasizes that the standard error of $\kappa$ depends on margins, yet often treats “margins” as a generic difficulty knob. We tested all three types of single-parameter simplifications: using only the primary category of Reader 1, using the average of the primary categories of both readers, and using only $1 – p_e$. Across 360 grid points where $K$ ranges from 2 to 6, $p_e$ ranges from 0.36 to 0.90, true TVD ranges from 0 to 0.20, and differences fall in near-end, far-end, or split-between-two-categories configurations, these simplifications fall within $\pm10\%$ of the true value only 12% to 19% of the time; the median absolute deviation ranges from 30% to 50%, and in the worst case, the error reaches 4.5 times. In other words, substituting $C$ with a single-variable function of $p_e$ introduces errors far beyond mere decimal precision.

    IV. Folded Normal Correction for Zero-Difference Categories

    The second trap is more subtle. Naive linearization has a failure point: when the true proportions of two readers are exactly equal in a certain category, the sign for that category becomes zero, and thus the sampling fluctuation of that category is entirely erased. However, “true equality” does not imply “observed equality”—the observed difference in that category fluctuates around zero, and its absolute value is not zero; it contributes the variance of a folded normal distribution: $0.25 \times (1 – 2/\pi) \times$ the variance of that category. Rewriting $C$ according to this correction yields a closed-form expression depending solely on the complete margins of both readers. In verification against 54 grid points, each with 51,200 cases and 10,000 Monte Carlo resamples, the median relative error between the closed form and Monte Carlo is less than 1%, with a maximum of 2.5%; all errors fall within 5%.

    V. The True Nature of $C$: A Joint Function of Margin Shape and Difference Location

    After correction, the structure of $C$ becomes clear. It is neither a function of $p_e$ nor of $\bar{m}$; rather, it is a joint function of “margin shape + difference location.” Holding $p_e$ and true TVD constant while varying only the location of the differences, the span of $C$ reaches 1.32 times when $K=3$ and 1.19 times when $K=6$. The example where $p_e=0.65$, $K=5$, and TVD=0.10 is particularly illustrative: when differences fall entirely between one primary category and one terminal category, $C$ is 0.2346, yielding a threshold of approximately 2,900 cases; with the same margins and TVD but differences split equally between two categories, $C$ rises to 0.2823, yielding a threshold of 3,500 cases. Two studies with identical true disagreement levels differ by 20% in required sample size solely because the “disagreement concentrates on one line” versus “spreads across two lines.”

    Why do the five $\bar{m}$ readings in #1409 appear monotonically decreasing and so regular? Because those five points all share the same configuration—differences concentrate between the primary category and the secondary category, while margins simultaneously shift from flat to sharp. Monotonic decrease is the product of “fixed configuration, sweeping margins,” not an intrinsic law of $C$. Once the configuration changes, $C$ can even reverse under the same $p_e$. This explains the anomaly in the low-consistency region: 0.60 and 0.75 lie on the same level, not due to measurement noise, but because two mechanisms cancel each other at this intersection—flatter margins cause per-category proportions to approach, increasing the number of non-zero components in the sign vector and raising the upper bound of variance; simultaneously, smaller differences narrow the true fluctuation.

    VI. This Result is Subtractive for Practice

    The implication of this result for practice is subtractive rather than additive. First, the threshold can be calculated before recruitment, but this requires reporting the complete expected margin distribution for each of the two readers, rather than just a single $p_e$ or a $\kappa$ target value—this aligns with methodological literature requirements: when solving for sample size based on target precision, the complete margin must be specified; providing only the coefficient cannot determine $n$. Second, any route of “estimating $C$ via a small pilot sample and extrapolating” is invalid, because $C$’s sensitivity to configuration is independent of $n$, and pilot studies cannot provide configuration information. Third, if practice allows only a single number, that number should take the upper bound based on the most unfavorable configuration—in this grid, configurations splitting across two categories generally yield higher $C$; using this as a conservative estimate will not underestimate the threshold.

    Evidence and Limitations

    Verified (pure algebra and Monte Carlo, reproducible): The linearized variance of TVD depends solely on the two readers’ margins; the sign vector and covariance matrix can be explicitly written; zero-difference categories must be supplemented with $0.25 \times (1 – 2/\pi) \times$ variance term via folded normal correction, otherwise naive linearization systematically underestimates; as the number of zero-difference categories increases from 0 to 3, the underestimation median rises from 0% to 12.4%, with a worst-case of 17%; the corrected closed form shows median relative error of 0.9% and maximum 2.5% against simulation across 54 grid points at scale of 50,000 cases.

    Partially Verified: $C$’s sensitivity to “difference location” (spanning 1.14 to 1.32 times under same $p_e$ and TVD) was swept only across three-category configurations (far-end single, near-end single, split-between-two-categories), not exhaustively over the entire support set; the rule “two-category-split configuration yields higher $C$” holds consistently across nine grid points where $K=3$ to 6 and $p_e=0.40$ to 0.80, but cannot be proven for all configurations; thus it serves as empirical basis for conservative upper bound rather than a theorem.

    To Be Verified: In real judgment studies, the distribution of “difference location” has never been systematically reported—published studies generally report only $\kappa$, AC1, and 95% intervals, rarely the complete margins for each of the two readers, making it impossible to infer configuration; this defines the applicability boundary of this conclusion and serves as the entry point for next work. Additionally, whether the closed form of $C$ remains valid when $K > 6$ or structural zero cells exist (categories never used by either reader) was not tested in this round.

    L1 Network Status: The L1 edge table and corresponding network file lack the edge “variance structure of judgment consistency metrics.” The closest is the E-060 label premise (status conflict: PMID 30383398 reports inter-physician diagnostic consistency in open populations as low; authors call it “concerning”; KB-317/334 reports AC2 has run successfully in real clinical scenarios, pointing out the practical pitfall that “prevalence $\neq$ assessor-observed prevalence”; KB-333 reports that the direction of $\kappa$’s sensitivity to ordinal/nominal scales opposes the Gwet series). This edge only registers “whether label reliability is sufficient,” not “what determines the variance of reliability metrics.” Hence this paper caps $C$ and does not advocate any verified mechanism. This topic has not yet entered the L1 corresponding network; the gap is registered.

    Further Reading

    Evidence Level: C
    References: PMID 11512137, PMID 22560852, PMID 36123642, PMID 38283651, PMID 23627889, PMID 25168681, PMID 15733050


    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.

  • 漂移和抽样误差不是两笔账,是同一笔数了两遍:那个相关系数不是 0.2,是 0.98

    Evidence level: C

    把漂移当作独立误差去和抽样误差相加,等于把同一块方差记了两次账;真正被漏掉的那一项不是协方差,是合格率对漂移的凹性。

    一、相加之前先问一句:这两项是两个东西吗

    1405 把漂移地板 5.31 个百分点与 m̂ 的抽样误差按方差相加,n=60 得 6.01 个百分点。这个做法只有一个前提:两项独立。但这两项的来源根本分不开——漂移是判读者的位置在变,m̂ 是同一批判读者在变过位置的尺子上判出来的合格比例。漂移不是 m̂ 之外的另一股噪声,漂移就是 m̂ 的生成过程。 一个量怎么可能是自己生成过程的独立误差项。

    二、一阶上不是「有点相关」,是几乎完全相关

    把漂移 b 设成均值零、标准差 0.167 的正态扰动,m̂ 是按切点算出的合格率,两者的一阶相关可以直接积出来。实测:m=0.90 时是 0.9930,m=0.95 时是 0.9825,m=0.70 时是 0.9982。在全部条件上都超过 0.98。

    换算成那条待验证项想量的比值更直观:2 倍协方差除以漂移方差,在 m=0.95、σ=0.167 时等于 18.30,而解析预言是 2 除以切点处的标准正态密度,即 19.39。不是「多加两成」,是那一项本身被放大了十八倍。 原设问的判据是「相关使合计方差比独立假设大百分之二十」,可这个量根本没停在百分之二十这个量级上。

    三、所以那个 2Cov 不是修正项,是重复计入

    到这里会想:那就把 2Cov 加进去。不行,加进去就错得更远。原因是全方差公式已经把两块算全了:

    Var(m̂) 等于 E[m(1−m)/n] 加上 Var(m)。

    前一项是「给定尺子位置、数 n 个人」的二项抖动,后一项是「尺子位置本身在变」的全部后果。两项相加即完整方差,没有第三项可加。所谓 2Cov,是把这个分解误当成「两个独立误差相加」才冒出来的影子项,它和 Var(m) 是同一块方差的两种写法。原设问的真正问题也在这里:它把符号问题(加还是减)当成了关键,而实际是记账单位问题——这笔方差只能记一次。

    四、漏掉的那一项:合格率对漂移是凹的

    真正的偏差在别处。二项项里那个 m(1−m) 不是常数,它随漂移变化,而且是凹函数,所以它的期望大于在均值处取值——这是詹森不等式。实测比值:σ=0.167、m=0.95 时是 1.0374,即二项项被抬高 3.7%;σ=0.515 时是 1.3018,抬高 30.2%,且这个比例不随 n 变。 后者意味着即使把 n 抬到无穷,二项项也只按三成折扣计入。加上地板一起算,σ=0.167 时 n=60 的有效样本量只剩名义的 41.5。

    方向也值得记一笔:E[m] 本身被漂移单向压低。m=0.95、σ=0.515 时算出 0.92818,缺口 −2.18 个百分点,是 5.31 这个一阶项的 41.1%。漂移对判据线的真实作用不是让 m̂ 上下抖,而是把合格率的期望值单向拽下去。

    五、两篇的口径其实对不上

    还有一处不自洽要摊开。5.31 个百分点反解出的漂移标准差是 0.5149,是 #1404 用值 0.167 的 3.08 倍。同一个 0.167 标准差在不同切点上换算出的百分点差着 14.8 倍:m=0.50 时 6.66,m=0.95 时 1.72,m=0.99 时 0.45。漂移的百分点值不可跨研究移植,只报百分点的漂移量没有可比性。

    后果落到地板上:按 5.31 反推的真地板是 6.97,不是 5.31。n=60 时真合计 7.67 对 6.01,低估 27.6%;n 抬到 2000 时真值仍有 6.99,那个 5.33 的收敛值本身就是错的。换回 0.167 那一档,要把标准误压到 2.00,#1405 式算出 n=119 就够,按真公式要 466——四倍差。

    六、四条判据

    一、记账单位:凡报漂移与抽样的合计误差,只许用全方差公式,禁写成「漂移方差 + 抽样方差 + 2Cov」。

    二、设计效应必须并报:须报 k 等于 E[m(1−m)] 除以 m0(1−m0),并由 k 反算有效样本量;只报名义 n 者其精度不可用于设计。

    三、漂移以标准差报:或同时报切点处的密度;只报百分点者禁作跨研究比较。

    四、地板先行:先算漂移导致的地板,目标标准误低于地板者判不可达,不得以加样本作答。

    证据与局限

    已验证:全方差公式、一阶相关 0.98 以上、2Cov 与 Var(m) 之比约 18 倍、凹性比值 1.0374 与 1.3018、詹森缺口 41.1%、n=119 与 466 的对照——全部是解析式加数值积分,可复算到任意精度,不依赖实测数据。

    部分验证:漂移取正态、判读取单切点阈值模型,这是最简设定;真实判读有多档、有读者特异成分,k 会偏离 1.30 这个数。

    待验证:① 多档判读下 k 实际上界;② 漂移含读者特异成分时一阶相关掉到多少(共同模与差模混在一起时,m̂ 只吃共同模);③ 切点也随漂移移动时,凹性与一阶项是否互相抵消。

    该主题尚未进入 L1 对应网络:研究内核的边表止于 E-061,网络正文第 202 行只有 E-060 标签前提这一条与一致性有关,且只裁决「标签信度低不低、训练能否改善」。全表检索漂移、协方差、方差成分、全方差、设计效应、凹性、詹森均无对应边。

    L1 网络状态:无对应边(E-060 标签前提仅登记「一致性低」与「训练可改善」的对立,状态 conflict 未裁决;网络无「误差结构」「方差记账」「设计效应」任何节点)

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 2348207, PMID 15939215, PMID 26925540, PMID 23343209, PMID 29881127, PMID 34893998, PMID 31018678


    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。

    Drift and sampling error are not two separate accounts but the same number counted twice: that correlation coefficient is not 0.2, it is 0.98.

    Evidence level: C

    Treating drift as an independent error to be added to sampling error equals double-counting the same variance component; the term truly omitted is not the covariance, but the concavity of pass rate with respect to drift.

    I. Before Adding, Ask: Are These Two Items Distinct Entities?

    1405 sums the drift floor of 5.31 percentage points with the sampling error of m̂ (n=60) via variance addition to obtain 6.01 percentage points. This approach rests on a single prerequisite: independence between the two terms. Yet their origins cannot be disentangled—drift arises from shifts in the rater’s position, while m̂ is the pass rate measured on that same shifting scale. Drift is not an additional noise source outside m̂; drift constitutes the generative process of m̂ itself. How can a quantity be an independent error term of its own generative process?

    II. At the First Order, It Is Not “Somewhat Correlated,” It Is Nearly Perfectly Correlated

    Setting drift b as a normal perturbation with mean zero and standard deviation 0.167, and m̂ as the pass rate calculated at the cutpoint, their first-order correlation can be directly computed. Measured values: when m=0.90 the correlation is 0.9930; when m=0.95 it is 0.9825; when m=0.70 it is 0.9982. Under all conditions, the correlation exceeds 0.98.

    Expressing this in terms of the ratio of the term to be verified becomes more intuitive: twice the covariance divided by the drift variance equals 18.30 when m=0.95 and σ=0.167, whereas the analytical prediction is 2 divided by the standard normal density at the cutpoint, yielding 19.39. It is not “adding two-tenths more”; that term itself is amplified eighteenfold. The original criterion asked whether “correlation increases the total variance by twenty percent,” but this quantity never reaches that order of magnitude.

    III. Therefore That 2Cov Term Is Not a Correction; It Is a Double Counting

    At this point one might think: then add 2Cov. No; adding it makes the error worse. The reason is that the full variance formula already accounts for both components:

    Var(m̂) equals E[m(1−m)/n] plus Var(m).

    The first term represents the binomial jitter given a fixed ruler position and counting n individuals; the second term captures all consequences of the ruler position itself shifting. Their sum constitutes the complete variance, leaving no third term to add. The so-called 2Cov is merely a shadow term arising from misinterpreting this decomposition as “two independent errors added together.” It and Var(m) are simply two expressions of the same variance component. The real issue in the original question lies here: it mistook the sign problem (plus or minus) for the critical factor, whereas the actual issue is one of accounting units—this variance can be counted only once.

    IV. The Omitted Term: Pass Rate Is Concave with Respect to Drift

    The true discrepancy lies elsewhere. Within the binomial term, m(1−m) is not constant; it varies with drift and forms a concave function, so its expectation exceeds the value at the mean—this is Jensen’s inequality. Measured ratios: when σ=0.167 and m=0.95 the ratio is 1.0374, meaning the binomial term is elevated by 3.7%; when σ=0.515 it becomes 1.3018, elevated by 30.2%, and this proportion does not vary with n. The latter implies that even if n is raised to infinity, the binomial term contributes only at a thirty percent discount. Including the floor together, when σ=0.167 and n=60, the effective sample size drops to only 41.5 of the nominal value.

    The direction also warrants note: E[m] itself is unidirectionally depressed by drift. When m=0.95 and σ=0.515, the calculated value is 0.92818, leaving a gap of −2.18 percentage points, which equals 41.1% of the first-order term 5.31. The true effect of drift on the criterion line is not to make m̂ jitter up and down, but to unidirectionally pull the expected pass rate downward.

    V. The Two Papers’ Calibers Are Inconsistent

    One further inconsistency must be laid bare. The drift standard deviation reverse-engineered from 5.31 percentage points is 0.5149, which is 3.08 times the value of 0.167 used in #1404. The same standard deviation of 0.167 yields percentage differences varying by a factor of 14.8 across different cutpoints: 6.66 when m=0.50, 1.72 when m=0.95, and 0.45 when m=0.99. The percentage-point value of drift cannot be transferred across studies; reporting only percentage-point drift renders comparisons meaningless.

    Consequences fall onto the floor: the true floor reverse-calculated from 5.31 is 6.97, not 5.31. When n=60, the true sum is 7.67 versus 6.01, underestimating by 27.6%; when n is raised to 2000, the true value remains 6.99, meaning that convergence value of 5.33 is itself wrong. Returning to the 0.167 tier, to suppress the standard error to 2.00, #1405’s formula yields n=119 as sufficient, whereas the true formula requires 466—a fourfold difference.

    VI. Four Criteria

    I. Accounting Units: When reporting the combined error of drift and sampling, use only the full variance formula; forbid writing “drift variance + sampling variance + 2Cov.”

    II. Design Effect Must Be Reported: Report k equal to E[m(1−m)] divided by m0(1−m0), and derive the effective sample size from k; reporting only the nominal n renders the precision unusable for design.

    III. Drift Reported as Standard Deviation: Or simultaneously report the density at the cutpoint; reporting only percentage points prohibits cross-study comparison.

    IV. Floor First: Calculate the floor caused by drift first; if the target standard error is below the floor, deem it unachievable and do not answer by adding samples.

    Evidence and Limitations

    Verified: Full variance formula, first-order correlation above 0.98, ratio of 2Cov to Var(m) approximately 18 times, concavity ratios 1.0374 and 1.3018, Jensen gap 41.1%, comparison of n=119 and 466—all derived from analytical expressions plus numerical integration, reproducible to arbitrary precision without relying on empirical data.

    Partially Verified: Drift modeled as normal, rater judgment using a single cutpoint threshold model—this is the simplest setup; real judgments involve multiple tiers and reader-specific components, causing k to deviate from 1.30.

    Pending Verification: ① the actual upper bound of k under multi-tier judgment; ② how much the first-order correlation drops when drift contains reader-specific components (when common mode and difference mode mix, m̂ consumes only the common mode); ③ whether concavity and the first-order term cancel each other when cutpoints also shift with drift.

    This topic has not yet entered the L1 corresponding network: the edge table of the research core stops at E-061; in the network body at line 202 there is only the E-060 label premise concerning consistency, and it merely adjudicates “whether label reliability is low and whether training can improve.” A full table search for drift, covariance, variance components, full variance, design effect, concavity, and Jensen yields no corresponding edges.

    L1 Network Status: No corresponding edges (the E-060 label premise only registers the opposition between “low consistency” and “training can improve,” with the conflict state unresolved; the network lacks any nodes for “error structure,” “variance accounting,” or “design effect”)

    Related Reading

    Evidence Level: C
    References: PMID 8501467, PMID 2348207, PMID 15939215, PMID 26925540, PMID 23343209, PMID 29881127, PMID 34893998, PMID 31018678


    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.

  • 两人一起偏和两人偏得不一样,到底谁在替谁背锅:共同偏移这条轴被三个不同的东西同时冒充

    Evidence level: C

    代数上 E 与 Δ 正交,所以 r 被消掉;但正交只是坐标变换的产物。真去测的时候,共同偏移这条轴根本不由「一起偏」生成,而由两人精度不同生成——投影出来的东西是假的。

    一、那道代数恒等式是真的,但它什么都没证明

    上一轮(#1407)推出一个漂亮的式子:判据的放宽倍数可写成 1+2E,其中 E 是两位判读者的共同偏移、Δ 是两人的偏离差。推导里有一句关键话:E 与 Δ 在代数上正交,所以两人的偏差和 r 在式子里消掉了,倍数只由 E 决定。这在数学上没错,但它只说明两件事能以这种坐标写下来,不代表这两个坐标在数据里能被分别看见。

    所有把这些论文放在一起看的读者都会撞上同一个坑。医学统计早就区分系统差与随机差,也早知偏倚和患病率会一起把 kappa 压下去(PMID 7701145、PMID 8501467)。但区分的前提是两者的生成机制不同——这不是观测结论,是建模假设。假设成立,分解才有意义;不成立,分解出的两个数只是同一堆噪声的两种切法。本篇就是把这条假设拿去撞一遍。

    二、把三种「一起偏」分别投到同一对坐标上

    做法很直接:造一个已知真值的判读场景,两位判读者对同一批材料各自独立落档,再计算 E 与 Δ 的样本相关。真值设为无任何关联,那么测出来的相关全是伪的。四种机制分开注入,看哪一种真能把 E 和 Δ 搅在一起。

    结果很干脆:共用的材料难度项(两人被同一批材料同时推移)无论强度从 0 加到 2.5,测出的相关一直在正负 0.015 以内,等于零。共同平移本身不生事。真正制造相关的只有两个东西:两人精度不等,噪声比 2 倍时有 0.455、3 倍时有 0.628;两人敏感度斜率不等,斜率比 2 倍时给 −0.241,符号稳定。换成可读的话:所谓「两人一起往同一边偏」,在数据里根本不是共同偏移起作用,是其中一人比另一人更不稳。

    三、敏感度这项得靠共用难度才发作,单独注入等于零

    上一条还藏着一个更细的结构。斜率不等单独注入、不给共用材料难度时,相关只有 −0.0013,测不出东西。把共用难度逐步加大,同样的斜率比 2.0 才依次走到 0.068、0.145、0.163、0.123——先涨后落。斜率不等是交互项,要借共同难度这个载体才显形。

    这一条把 backlog 原来的假设整个翻过来。原来担心「共同难度让两人一起偏,于是 E 与 Δ 混在一起」;实测是共同难度单独完全不生事,它是斜率不等与精度不等这两个真凶的放大器。而斜率比从 1 到 2,相关从 0 走到 −0.482,符号还相反,与精度不等正对着干。两个真凶在同一份数据里互抵,谁大谁说话。

    四、按难度分层不是洗掉它,是把假的换成另一个假的

    backlog 给的对策是按材料难度分层,层内重算,看相关是否消失。这条对策实测下来是加害者。

    在共用难度很强(2.5)、斜率相等(真值应为零)的情形下,不分层的相关是 0.0037,几乎干净;按真实难度三分位分层后,层内平均变成 −0.004,还是干净。但只要斜率比 1.6,不分层只 0.094,分层后跳到 −0.325;斜率比 2.0 时从 0.128 跳到 −0.430。精度不等的情形同样:噪声比 2 倍,不分层 0.017,分层后 0.371;3 倍时从 0.021 跳到 0.432。

    根子在分层这个动作本身。条件于共用难度之后,两人潜在读数仍只靠各自噪声分开,而把连续值切进档位这一步是非线性的,会把「同层内一人偏高必然另一人偏低」的互补关系放大成负相关。层内两读者原始档位的相关,从 0.200 一路掉到 0.033、0.133、0.221、0.326——分层剥掉的正是两人真正共有的部分,剩下的是切档放大的人工互补。分层把「共同难度造成的伪相关」换成了「切档非线性造成的伪相关」,而后者更大。

    五、三档量表自己就有一个天花板,0.6 这条线根本到不了

    backlog 给的判据是:相关绝对值小于 0.2 则两轴可分立;大于 0.6 则不可分、1+2E 只能做事后诊断。这两个数在真实的三档量表上都不可能实现。

    把参数全扫一遍(噪声比 1 到 5、共用难度 0 到 3、斜率比 1 到 2),三档数据能达到的相关绝对值上限只有 0.391。原因在分辨率:档位越粗,E 与 Δ 的信息被压得越扁。二档能到 0.873,三档掉到 0.205,五档 0.612,十一档 0.705。所以「0.6 以上判不可分」对三档材料是永不可能触发的死条款,「0.2 以下判可分立」则会把真值 0.39 误判成可分立。

    估计精度更麻烦。n=120 时单次相关的标准差 0.0785,真值为零也有约 1.5% 的实验读出超过 0.2;要看穿 0.2 这个量级,n=400 才把检出率做到 71.5%,n=800 才 97.2%。而 1+2E 这个修正项,n=120 时 90% 区间宽 0.250,比它要修正的量还大。

    六、二十人时这条轴的性质变了,不能从两人情形外推

    backlog 的实验设计是 20 位判读者。这个规模下 E 与 Δ 的定义要重写:E 变成二十人的平均偏移,Δ 变成二百一十对差值的平均。等精度等斜率时相关是 −0.003,干净。但只要一人斜率 1.6,相关变 0.070;十人精度差两倍,变 −0.244。方向与两人情形也不一致——两人时精度不等给正、二十人时给负。

    所以两人情形的结论不能直接搬到多人设计上。而修正项的可估性在多人时略有改善,却仍不足:n=480 时 2E 的 90% 区间宽 0.133,n=1920 时才降到 0.067。要让修正项窄于它要修正的量,材料得过千例,这在针灸得气这类现场判读里不现实。

    证据与局限

    已验证: 共用加性难度项不产生 E 与 Δ 的相关,这是模拟里的稳定结论(五种强度全部落在 ±0.015 内),且有代数根据。精度不等与斜率不等确实产生非零相关,方向确定且随参数单调。三档量表的相关系数存在分辨率天花板(0.391),随档位增加而抬升,这是离散化的结构性质。按共同难度分层会诱导更大的伪相关,方向与幅度在两组参数下一致。

    部分验证: 分层诱导伪相关的机制解释(条件化后切档放大互补)在模拟里方向明确,但没有在真实判读数据上验证过。二十人与两人情形的方向不一致只在三组参数下出现,样本不足以断言规律。1+2E 修正项的精度随例数改善的速度是按模型算的,真实判读的误差结构更复杂。

    待验证: 真实判读材料里三种不对称的实际混合比例未知,本篇只能给出「哪种机制给多大相关」,给不出「实际数据里各占多少」。贴阈例子的占比是否真如假设那样影响结论,需要真实材料检验。斜率不等与精度不等在真人身上是否独立、还是一起变化,本篇当作独立注入,这一条没测。

    相关阅读

    证据等级: C
    参考文献: PMID 7701145, PMID 8501467, PMID 26095449, PMID 30080056, PMID 25090041, PMID 33853549
    L1 网络状态: 该主题尚未进入 L1 对应网络。最接近的边是 E-060 标签前提(五脏辨证作为分组标签|前提|标签信度),该边在 L1 中标为「冲突(未裁决)」,涉及 KB-292/KB-317/KB-333/KB-334 关于中医师间诊断一致性的争论;本篇讨论的「共同偏移与偏离差能否分别报告」属于该边的估计层前提,但 L1 中尚无对应条目(检索「共同偏移」「偏离差」「方差成分」「GLMM」「可辨识」均为 0 命中)。


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

    When Two People Deviate Together, Their Deviations Differ: Who Is Taking the Blame for Whom? Three Distinct Things Are Simultaneously Impersonating This Shared Axis of Deviation.

    Evidence level: C

    Algebraically, E is orthogonal to Δ, so r is eliminated; yet orthogonality is merely a product of coordinate transformation. In true measurement, the axis of common offset does not arise from “drifting together,” but from differing precision between the two observers—the projected result is spurious.

    I. That algebraic identity holds, but proves nothing

    In the previous round (#1407), a neat formula was derived: the relaxation factor of the criterion can be written as 1+2E, where E represents the common offset shared by two readers and Δ denotes their deviation difference. A key statement in the derivation notes that E and Δ are algebraically orthogonal, causing the sum of deviations r to vanish from the equation, leaving the multiplier determined solely by E. This is mathematically correct, yet it only indicates that two quantities can be expressed in this coordinate system; it does not mean these coordinates can be separately observed in the data.

    Readers who assemble all these papers will inevitably encounter the same pitfall. Medical statistics long ago distinguished systematic error from random error, and also knew that bias and prevalence jointly depress kappa (PMID 7701145, PMID 8501467). However, the premise of this distinction is that their generation mechanisms differ—this is not an observational conclusion but a modeling assumption. If the assumption holds, decomposition is meaningful; if not, the two decomposed numbers are merely two ways of slicing the same noise. This paper tests that assumption head-on.

    II. Projecting three types of “drifting together” onto the same pair of coordinates

    The approach is straightforward: construct a reading scenario with known truth values, have two independent readers classify the same batch of materials, then compute the sample correlation between E and Δ. If truth values are set to have no association, any measured correlation is entirely spurious. Each of the four mechanisms is injected separately to see which one truly mixes E and Δ.

    The results are crisp: the shared difficulty component (both observers shifted by the same batch of materials) yields a correlation within ±0.015 regardless of intensity ranging from 0 to 2.5, effectively zero. Common translation itself generates no correlation. Only two factors create correlation: unequal precision between observers (with noise ratio 2, correlation is 0.455; with noise ratio 3, it is 0.628); unequal sensitivity slopes (with slope ratio 2, correlation is −0.241, with stable sign). In plain terms: the notion of “both drifting together in the same direction” does not reflect common offset acting in the data; rather, it reflects that one observer is less stable than the other.

    III. Sensitivity requires shared difficulty to activate; injecting it alone yields zero

    The previous section also conceals a finer structure. When slope inequality is injected alone without shared difficulty, correlation is only −0.0013, yielding no measurable effect. As shared difficulty increases stepwise, the same slope ratio of 2.0 sequentially produces correlations of 0.068, 0.145, 0.163, 0.123—rising then falling. Slope inequality acts as an interaction term; it becomes visible only through the carrier of shared difficulty.

    This finding overturns the original backlog assumption entirely. Previously, concern focused on “shared difficulty causing both to drift together, thus mixing E and Δ.” In reality, shared difficulty alone does nothing; it amplifies the two true culprits: unequal slopes and unequal precision. Meanwhile, as slope ratio increases from 1 to 2, correlation moves from 0 to −0.482, with the sign opposite to that of unequal precision, directly opposing it. In the same dataset, the two culprits counteract each other; whichever is stronger dominates.

    IV. Stratifying by difficulty does not wash it away—it replaces one falsehood with another

    The countermeasure proposed in the backlog is to stratify by material difficulty, recalculate within each layer, and observe whether correlation disappears. In practice, this strategy proves harmful.

    Under strong shared difficulty (2.5) and equal slopes (truth should be zero), the unstratified correlation is 0.0037, nearly clean; stratifying by true difficulty tertiles keeps the within-layer average at −0.004, still clean. However, with slope ratio 1.6, unstratified correlation is only 0.094, whereas stratification jumps it to −0.325; with slope ratio 2.0, it leaps from 0.128 to −0.430. The same pattern holds for unequal precision: with noise ratio 2, unstratified correlation is 0.017, stratified becomes 0.371; with noise ratio 3, it rises from 0.021 to 0.432.

    The root lies in the stratification action itself. Conditional on shared difficulty, the two potential readings still separate only via their respective noise. Cutting continuous values into bins is nonlinear, amplifying the complementary relationship—”one higher within the same layer necessarily implies the other lower”—into negative correlation. The raw within-layer correlations of the two readers drop from 0.200 to 0.033, 0.133, 0.221, 0.326—stratification strips away precisely the part they truly share, leaving behind artificial complementarity magnified by binning. Stratification replaces “spurious correlation caused by shared difficulty” with “spurious correlation caused by nonlinear binning,” and the latter is larger.

    V. The three-tier scale itself has a ceiling; the 0.6 threshold is never reached

    The backlog criterion states: if absolute correlation < 0.2, the two axes can be separated; if > 0.6, they cannot—and 1+2E serves only as post-hoc diagnosis. These thresholds are unattainable on real three-tier scales.

    Scanning all parameters (noise ratio 1–5, shared difficulty 0–3, slope ratio 1–2), the maximum absolute correlation achievable on three-tier data is only 0.391. The reason lies in resolution: coarser tiers compress the information of E and Δ more severely. Two-tier reaches 0.873, three-tier drops to 0.205, five-tier hits 0.612, eleven-tier achieves 0.705. Thus, “unseparable if >0.6” is a dead clause never triggered for three-tier materials; “separable if <0.2” would falsely classify true value 0.39 as separable.

    Estimating precision is even more problematic. With n=120, the standard deviation of single correlation is 0.0785; even with true zero, about 1.5% of experimental readings exceed 0.2. To detect the 0.2 scale, n=400 yields 71.5% detection rate, n=800 reaches 97.2%. Meanwhile, the correction term 1+2E: with n=120, its 90% interval width is 0.250, larger than the quantity it aims to correct.

    VI. With twenty observers, the nature of this axis changes; conclusions from two cannot be extrapolated

    The backlog experimental design involves 20 readers. At this scale, the definitions of E and Δ must be rewritten: E becomes the average offset across twenty people, while Δ becomes the average of 210 pairwise differences. Under equal precision and equal slopes, correlation is −0.003, clean. But if one person has slope 1.6, correlation shifts to 0.070; with ten people having twice the precision difference, it becomes −0.244. The direction differs from the two-person case—unequal precision yields positive correlation for two, negative for twenty.

    Thus, conclusions from the two-person scenario cannot be directly transferred to multi-person designs. While the estimability of the correction term improves slightly with more people, it remains insufficient: with n=480, the 90% interval width of 2E is 0.133; only at n=1920 does it drop to 0.067. To make the correction term narrower than the quantity it must correct, materials would need over a thousand cases—impractical in real-world reading tasks like acupuncture qi sensation.

    Evidence and Limitations

    Verified: Shared additive difficulty does not generate correlation between E and Δ, a stable conclusion from simulation (all five intensities fell within ±0.015), supported by algebraic reasoning. Unequal precision and unequal slopes indeed produce non-zero correlation, with determined direction and monotonic dependence on parameters. The correlation coefficient for three-tier scales has a resolution ceiling (0.391), rising with tier count, reflecting the structural nature of discretization. Stratifying by shared difficulty induces larger spurious correlation, with direction and magnitude consistent across both parameter sets.

    Partially Verified: The mechanism explanation for stratification-induced spurious correlation (amplification of complementarity after conditioning and binning) is directionally clear in simulation but has not been validated on real reading data. The directional inconsistency between twenty-person and two-person cases appears only under three parameter sets, insufficient to establish a general rule. The precision improvement rate of the 1+2E correction term with sample size follows model calculation, while real reading error structures are more complex.

    To Be Verified: The actual mixing ratios of the three asymmetries in real reading materials remain unknown; this paper can only indicate “which mechanism produces how much correlation,” not “what proportion each occupies in real data.” Whether threshold-proximate cases truly affect conclusions as assumed requires validation with real materials. Whether slope inequality and precision inequality operate independently or co-vary in humans has not been tested; this paper assumes independent injection.

    Related Reading

    Evidence Level: C
    References: PMID 7701145, PMID 8501467, PMID 26095449, PMID 30080056, PMID 25090041, PMID 33853549
    L1 Network Status: This topic has not yet entered the L1 corresponding network. The closest edge is the E-060 label premise (organ differentiation as grouping label | premise | label reliability), which is marked as “Conflict (Unresolved)” in L1, involving debates among KB-292/KB-317/KB-333/KB-334 regarding diagnostic consistency among TCM practitioners; the issue of “whether common offset and deviation difference can be separately reported” discussed here belongs to the estimation layer of that edge, but no corresponding entry exists in L1 (searches for “common offset,” “deviation difference,” “variance component,” “GLMM,” “identifiable” all yield zero hits).


    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.

  • 想从别人的论文里倒推出那个门槛,倒不出来:他们判「合格」用的根本不是这把尺子

    Evidence level: C

    反推一个参数的前提,是被判定的那个量真的由这个参数决定。而文献里「一致性可接受」这句话,是由另一套指标决定的。

    一、这个想法本身很干净

    上一轮的缺口是:决定「差多少算不能容忍」的带宽 w 来源不明,0.05 与 0.15 都能自称可接受,宽出三倍。于是有了一条出路——既然 w 是客观存在、只是没人测过的量,就从已发表研究里倒推。每篇报告了双读者边际分布的研究都隐含地提供了一个 w:作者写下「一致性可接受」时,心里必在用某个默认带宽。把这些 w 按结论分组,分离得开则经验值就出来了。这推理无懈可击,但跳过了一句前提。

    二、被判定的量必须由那个参数决定

    反推要成立,作者判「可接受」时用的判据得真是 w 那条线。这听着像同义反复,实际不是。判读研究报的指标主流为三类:观察一致率、kappa 一类偶然校正系数、组内相关系数,阈值长期由教科书建议给定。系统综述给了直接证据:七十九篇观察者变异研究中五成二报组内相关系数、三成九报 kappa,而作者指出这些变异估计的解读常与报告结论不对应(PMID 37399082)。结论那句话与它前面那个数字之间,本就有一处没人核对的缝。

    更麻烦的是两套指标测的不是同一件事。kappa 校正全部不一致,涵盖三档之间所有错配;带宽只问主档占比差了多少,两者没有常数换算。用阈值潜变量模型仿真(三档、双读者、每批三十到五百例、噪声 0.3 至 0.9 个标准差,真判据设为 kappa 不低于 0.60):用主档占比差去分割「可接受」与「不可接受」两组,最优分割点最小误分类率 19.7%——五分之一的研究被放错组;门槛放宽到 kappa 0.40 则升到 42.0%,接近抛硬币。反过来若真判据本就在带宽空间,同样分割的误分类率是 0.0%。

    这不是方法钝,而是方向单向:真判据在带宽空间时反推完美,在 kappa 空间时近乎无效。两量相关系数只有负 0.105,几乎正交,而文献里的判据恰恰在后一个空间。

    三、第二层:噪声比要测的门槛还大

    就算空间摆平,还有更硬的障碍。每篇研究几十到一两百例,两位读者在同一批案子上的主档占比差本身就是随机量。仿真给出这个差的标准差:三十例 0.062、四十七例 0.051、八十八例 0.037、一百二十例 0.031。影像变异研究的病例数中位数正是四十七例(PMID 37399082)。两者一叠就清楚:四十七例时噪声正好是 w 取 0.05 的 1.0 倍、取 0.10 的 0.5 倍。一篇中位规模的研究,报出的差里一半以上是噪声。 拿这堆读数反推 0.05 门槛,等于用比被测物还粗的尺子去量。

    加篇数能救一部分。若判据真在带宽空间、真值 0.10,篇数由三十加到六百,反推的九十五分位从 0.070 收敛到 0.074,区间由正负 0.018 收窄到 0.005。但它收敛到 0.074 而非 0.100,偏低 26%——被接受那组截在门槛下,用组内分位数估门槛必然偏小;且这收敛只在最有利世界成立。

    四、双峰判不出来,但双峰是真的

    原题有一条判据:若反推的 w 明显双峰,说明「可接受」在文献里是两个互不相通的学派。这条数据上可能成立,方法上却不可达。

    设两个领域真各用一套门槛,一为 0.05 一为 0.15。各自被接受的研究读数仍带噪声(一百二十例时 0.031),两峰相距 0.10 是噪声 3.2 倍,听上去够分。关键在截断:被接受意味着读数落在门槛以下,两组都是右偏截尾分布,不是两个居中钟形。截尾改掉了峰的形状——0.05 那支被截成贴着零的窄条,0.15 那支是宽得多的斜坡,混起来是一条单调下降曲线,没有第二峰。

    实测一致:高斯混合做单峰对双峰比较,双峰世界检出率 100%,单峰截尾世界假阳性率同样 100%。两种世界的读数长得一模一样。 不是模型不够强,是两种生成过程在截尾后信息等价。故「看分布有几个峰」这条判据,无论结论是双峰还是单峰都不作证据。

    五、结论

    第一,从已发表一致性研究反推 w,这条方法上关闭。不是篇数不够,是被判定的量与待反推的量不在同一空间,用带宽分割 kappa 判定的组,误分类率 19.7% 起。任何声称「从两百篇文献反推出 w」的工作,都得先证明那两百篇的判据真在带宽空间——按现有报告规范,这证明做不出来。

    第二,「反推不出统一门槛,所以法则的实证基础不存在」也不成立。原题设了退路:若 w 与结论分离度接近零,说明裁决没用可量化阈值。但分离度接近零有两个原因——裁决真没有门槛,或门槛设在另一个坐标系。本篇给出了后者的量级:在 kappa 世界分离度就是接近零,而那里的门槛明确得不能再明确。把「测不出」读成「不存在」,是这类反推最容易踩的坑。

    第三,能救这条路的只有判据写清楚的数据。判据一旦写成「主档占比差不超过百分之几」,反推就精确;反之只报 kappa 与置信区间、不报边际分布的研究,多少篇都不进池子。这压力转回「改报告规范」——既有共识文件早已要求:可靠性研究报告长期不完整,样本选择、设计、统计信息常缺失,使结果解读与合并困难(PMID 21130355)。

    证据与局限

    已验证: 阈值潜变量模型下,判据位于带宽空间时反推精确(误分类率 0.0%)、位于 kappa 空间时反推失效(19.7% 至 42.0%),两空间相关系数负 0.105。本轮仿真直接结果,可复现。

    部分验证: 影像观察者变异研究的方法学现状——七十九篇中病例数中位四十七例、观察者四人、仅 15% 论证样本量、五成二报组内相关系数、三成九报 kappa、且「变异估计的解读常与结论不对应」(PMID 37399082)。为判据空间错配提供现实基础,但该综述未逐篇核对判据口径,只作间接支撑。

    待验证: ① 真实文献中判据写成带宽形式的比例,需逐篇抽取,若低于一成则反推彻底无料;② 截尾导致的单峰与双峰信息等价,是否在真实分层数据上同样表现为「怎么分都看不出两峰」;③ 反推估计量系统性偏低 26% 能否用截尾反卷积校正回来。

    证据等级: C
    参考文献: PMID 37399082, PMID 21130355, PMID 15733050, PMID 2348207, PMID 2189948

    相关阅读

    L1 网络状态:L1 中最近的一条边为 E-060 标签前提(五脏辨证作分组标签|前提|标签信度,状态 conflict):「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)」,冲突另一方为 KB-317:「抗灰区最优的是 Gwet AC2 与 Brennan-Prediger κ(PMID 33853549 蒙特卡洛),但所有加权系数在『中低一致性 + 大样本 + 类别数多』时精度显著下降」。该边只登记「标签信度够不够用」,未登记「判定用的指标与待反推的带宽参数是否同一坐标系」。L1 网络中「kappa」「反推」「最小重要差」「判读」「w」命中均为 0,「一致性」6 处与「信度」8 处全部落在 E-060,「可接受」1 处为内感受量表因子结构性质,与判读门槛无关;边表 EDGE_TABLE 概念最高止于生理概念,无一致性度量节点。故本文按规则封顶 C,上述结论应读作「方法学层面的可辨识性论证」,不作机制断言。


    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。

    They didn’t use that ruler to judge \”qualified.\

    Evidence level: C

    A necessary premise for reverse-engineering a parameter is that the measured quantity is truly determined by that parameter. The phrase “acceptable consistency” found in literature is instead governed by a different set of metrics.

    I. The Core Idea Is Conceptually Clean

    The gap in the previous round was that the bandwidth $w$ defining “what difference is intolerable” had no known source; values of 0.05 and 0.15 both claimed to be acceptable, differing by a factor of three. This created an apparent impasse—until one considers that $w$ is an objective quantity simply never measured.

    This opens a path: since published studies reporting the marginal distribution of two readers implicitly provide a $w$, we can reverse-engineer it. When authors state “consistency is acceptable,” they are necessarily using some default bandwidth as their mental criterion. By grouping these $w$ values according to the study’s conclusion, we can separate acceptable from unacceptable cases and derive empirical values. This reasoning is logically sound, yet it skips a critical premise.

    II. The Measured Quantity Must Be Determined by That Parameter

    For reverse-engineering to hold, the criterion authors use to judge “acceptable” must truly be the $w$ threshold. This sounds tautological, but it is not.

    The dominant metrics reported in studies fall into three categories: observed agreement rates, kappa-type chance-corrected coefficients, and intraclass correlation coefficients (ICC). Thresholds for these have long been set by textbook recommendations. Systematic reviews provide direct evidence: among 79 studies of observer variation, 52% reported ICC and 39% reported kappa, yet authors noted that interpretations of these variation estimates often do not align with the reported conclusions (PMID 37399082). There is thus an unverified gap between the numerical result and the stated conclusion.

    Complicating matters further, these two metrics measure fundamentally different things. Kappa corrects for all inconsistencies across three categories, capturing every mismatch; bandwidth only asks how much the main category proportion differs. There is no constant conversion between them.

    Simulations using threshold latent variable models (three categories, two readers, 30–500 cases per batch, noise 0.3 to 0.9 standard deviations, true criterion set at kappa ≥ 0.60) reveal this disconnect:
    – When splitting “acceptable” vs. “unacceptable” based on main category proportion difference, the optimal cutoff yields a minimum misclassification rate of 19.7%—one-fifth of studies are misclassified.
    – If the threshold is relaxed to kappa ≥ 0.40, the error rate rises to 42.0%, nearly random.
    – Conversely, if the true criterion lies in bandwidth space, the same splitting yields 0.0% misclassification.

    This is not a methodological flaw; it reflects directional asymmetry: reverse-engineering works perfectly when the true criterion is in bandwidth space but fails almost entirely when it is in kappa space. The correlation coefficient between these two metrics is only −0.105, nearly orthogonal, yet the literature’s criteria operate precisely in the latter space.

    III. Second Layer: Noise Exceeds the Threshold

    Even if spatial alignment were resolved, a more fundamental obstacle remains. Each study involves dozens to hundreds of cases; the difference in main category proportions between two readers on the same batch is itself a random variable. Simulations yield standard deviations for this difference:
    – 30 cases: 0.062
    – 47 cases: 0.051
    – 88 cases: 0.037
    – 120 cases: 0.031

    The median case count in imaging variation studies is exactly 47 (PMID 37399082). Overlaying these reveals that at 47 cases, the noise equals $1.0 \times w$ if $w = 0.05$ and $0.5 \times w$ if $w = 0.10$. In a study of median size, more than half of the reported difference is pure noise.

    Increasing sample size helps partially. If the true criterion lies in bandwidth space with $w = 0.10$, expanding from 30 to 600 cases causes the reverse-engineered 95th percentile to converge from 0.070 to 0.074, narrowing the interval from ±0.018 to 0.005. However, it converges to 0.074 rather than 0.100—a 26% underestimate—because the accepted group is truncated below the threshold, and estimating the threshold via within-group quantiles inevitably underestimates it. Moreover, this convergence holds only in the most favorable scenario.

    IV. Bimodality Cannot Be Detected, Yet Bimodality Is Real

    The original criterion states: if the reverse-engineered $w$ shows clear bimodality, “acceptable” in the literature represents two distinct schools. This may hold empirically but is methodologically unreachable.

    Suppose two domains truly use different thresholds: $w = 0.05$ and $w = 0.15$. The accepted studies still carry noise (0.031 at 120 cases), and the two modes are separated by 0.10—3.2 times the noise—which seems sufficient. The critical issue is truncation: acceptance means the reading falls below the threshold, so both groups are right-truncated distributions, not centered bell curves. Truncation distorts the shape of the modes—the $w = 0.05$ mode becomes a narrow strip near zero, while the $w = 0.15$ mode is a much broader slope. Combined, they form a monotonically decreasing curve with no second peak.

    Empirical validation confirms this: Gaussian mixture modeling comparing single-mode vs. bimodal distributions yields 100% detection rate in the true bimodal world but also 100% false positive rate in the single-mode truncated world. The readings look identical in both worlds. It is not that the model is insufficient; rather, after truncation, the information in both generative processes becomes equivalent. Thus, “checking how many peaks the distribution has” provides no evidence, regardless of whether the conclusion is bimodal or unimodal.

    V. Conclusion

    First, reverse-engineering $w$ from published consistency studies is methodologically closed. It is not due to insufficient sample size, but because the measured quantity and the target parameter do not reside in the same space. Using bandwidth to split kappa-defined groups leads to misclassification rates starting at 19.7%. Any claim of deriving $w$ from hundreds of studies must first prove those studies’ criteria truly operate in bandwidth space—a proof impossible under current reporting standards.

    Second, “the inability to reverse-engineer a unified threshold implies no empirical basis for the rule” is false. The original setup allowed an escape: if $w$ and the conclusion separate completely, it suggests no quantifiable threshold exists. Yet this separation can arise from two causes: either the rule truly has no threshold, or the threshold lies in another coordinate system. This paper establishes the latter magnitude: separation is near zero in kappa space, where the threshold is unmistakably defined. Misreading “cannot be measured” as “does not exist” is the most common pitfall of such reverse-engineering.

    Third, the only way to salvage this approach is through explicitly stated criteria. If the criterion is written as “main category proportion difference not exceeding X percent,” reverse-engineering becomes precise. Conversely, studies reporting only kappa and confidence intervals without marginal distributions cannot contribute, regardless of number. This pressure shifts back to “improving reporting standards”—a consensus document already notes that reliability reports are often incomplete, lacking sample selection, design, and statistical details, making result interpretation and meta-analysis difficult (PMID 21130355).

    Evidence and Limitations

    Verified: Under threshold latent variable models, reverse-engineering is precise when the criterion lies in bandwidth space (misclassification rate 0.0%) but fails when in kappa space (19.7% to 42.0%). The correlation between the two spaces is −0.105. These simulation results are directly reproducible.

    Partially Verified: The current state of methodology in imaging observer variation studies—79 studies with median case count 47, four observers, only 15% arguing for sample size justification, 52% reporting ICC, 39% reporting kappa, and frequent misalignment between variation estimates and conclusions (PMID 37399082)—provides realistic foundation for criterion-space mismatch. However, this review did not verify criteria across all studies, offering only indirect support.

    To Be Verified:
    1. The proportion of true literature where criteria are expressed in bandwidth form must be extracted study-by-study; if below 10%, reverse-engineering yields nothing.
    2. Whether truncation-induced equivalence of single-mode and bimodal information holds in real stratified data, manifesting as “no detectable bimodality regardless of splitting method.”
    3. Whether the systematic 26% underestimation in reverse-engineered estimates can be corrected via deconvolution of truncation effects.

    Evidence Level: C
    References: PMID 37399082, PMID 21130355, PMID 15733050, PMID 2348207, PMID 2189948

    Related Reading

    L1 Network Status: The most recent edge in L1 is labeled E-060 (TCM diagnosis grouping label | premise | label credibility, status conflict): “KB-292 (conflict): Diagnostic consistency among TCM physicians in open populations is ‘low’ and described as ‘concerning’ (PMID 30383398); the second paper in the series notes that chance-corrected consistency statistics are often misused and recommends Gwet AC2 (PMID 31730403 / 31634000 / 31730404).” The conflicting side is KB-317: “The optimal gray zone is Gwet AC2 and Brennan-Prediger κ (PMID 33853549 Monte Carlo), but all weighted coefficients suffer significant accuracy loss under ‘low-moderate consistency + large sample + many categories’.” This edge only records “whether label credibility suffices,” omitting “whether the criterion metric and the reverse-engineered bandwidth parameter share the same coordinate system.” In the L1 network, “kappa,” “reverse-engineer,” “minimum important difference,” “judgment,” and “$w$” all hit zero; “consistency” appears in 6 places and “reliability” in 8 places, all falling under E-060, while “acceptable” appears once as an introspective scale factor unrelated to judgment thresholds. The EDGE_TABLE concept reaches only physiological concepts, lacking consistency measurement nodes. Therefore, this paper caps at C per rules; the above conclusions should be read as “methodological identifiability arguments,” not mechanistic assertions.


    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.

  • 记忆把 ICC 抬上去、求稳把它压回来:两个相反机制同时在场时,那个「0.5 附近」的数字什么都证明不了

    Evidence level: C

    ICC 是个比值。比值只报分子分母的比例,不报各自是多少——两个方向相反的机制同时作用时,同一个比值能对应六倍以上的真实差异。

    一、一个从没被正面检验的假设

    重测一致性的逻辑很朴素:同一人、同一批材料判两次,对得上说明尺子稳。这里藏着一个假设——两次对不上,只可能因为尺子在抖。

    真实判读里至少有两条路让两次不一样,方向却相反。第一条是记忆:第二遍看见同一批材料,判读者照着上遍的结论抄,抄一遍更一致,ICC 往上走。第二条是求稳:判读者意识到自己在重复同一件事,怕前后矛盾被挑出来,于是往中间档收敛。收敛也让打分接近,但那是把差异抹平,不是把信号锁住——ICC 不一定动,判别力却在掉。

    母篇已证明 ICC 对纯记忆敏感、对纯求稳近乎失明。这篇问下一层:两条路同时在跑时,报出来的那个 ICC 是哪一条的读数。答案是:哪条都不是,是两条相减的余数,而余数不能反推两个减数。

    二、ICC 的等值线是曲线,不是一个点

    把 ICC 锁在同一个数上看它下面压着什么最直接。实测:ICC 全部锁定 0.18 时,漂移跨度是 0.075 到 0.503,相差 6.7 倍。 无记忆时漂移 0.075 就够;记忆权重 0.6 时要 0.150;权重 0.9 时要 0.503。

    读法很刺眼。ICC 等于 0.18 的这个研究,可能是一位「没记忆、正常漂移」的判读者,也可能是「强记忆、漂移大到五倍」的判读者。前者可折价使用,后者必须整份作废——0.503 的漂移意味着第三轮和第一轮的尺度差出半个标准差,两轮不是同一把尺。而报告上只有「ICC=0.18」一个数字。

    四格检验更狠。按记忆强弱乘漂移大小分四组,各跑四百次取四分位区间:低记忆低漂移 0.121 到 0.227,低记忆高漂移 0.067 到 0.137,高记忆低漂移 0.131 到 0.234,高记忆高漂移 0.120 到 0.219。六对两两比对,六对全部重叠。 在这个常见规模下,「重测干净」和「重测已废」两群人的 ICC 分布分不开。

    三、加指标、加读者,都补不回来

    第一反应是「多报一个量」。本篇的补救指标直接测漂移、不走 ICC 那个比值,灵敏度确实碾压:权重 0.9 时漂移每涨 0.04,它的斜率是 ICC 斜率的三百五十六倍。但灵敏度不等于可辨识度。在 110 格网格上问「给定一组观测,容差内还有几格能解释」:只用 ICC,唯一可辨识率一律为零;双指标也只到 3.6%,平均每格还有 14.3 个候选解。把比值精度提到正负 4%(这个规模上基本做不到)才勉强到 28.2%。原因是 ICC 掉一点、漂移指标涨一点,这两个动作在中低记忆区几乎绑死——等于在本就压扁的曲面上多取一个几乎共线的投影。

    第二反应是「多找几位判读者」。实测也掐掉了:强记忆下把漂移从零拉到 0.20,ICC 位移是 4 位读者加 0.0068、8 位减 0.0119、16 位加 0.0011。位移一直在零附近晃,符号还翻了。 加读者只压小标准差(0.096 到 0.092,因为 ICC 的抽样误差主要由例数定),压不动那个本就接近零的位移。这不是精度问题是结构问题:两个反向项在同一条通道里,在同一通道上加观测也分不开。

    有个办法能分开,但要改设计:一半判读者判重复材料、另一半判同分布的新材料。重复臂走记忆、新材臂不走,两臂 ICC 之差就是记忆净效应,漂移在两臂相互抵消。实测方向对但信噪比极差——有记忆时两臂差 0.0522、无记忆 0.0104,要 32 到 48 位读者才能把单次功效值做到 0.30 到 0.37,按八成功效算要重复 57 到 87 次。原理成立,实践极贵。

    四、这对本组意味着什么

    本组所有跨主体比较都建立在同一前提上:分组标签与生理量是同一次测量得到的,判读重测一致性是这条链路唯一的质量闸门。现在这道闸门的样子是——它只报一个比值,而这个比值在两机制同向抵消时,与「一切正常」无法区分。

    三条后果。第一,凡报告「重测 ICC 达标故数据可用」的研究都缺一句声明:这个数是哪种机制主导下拿到的。记忆主导说明重测根本没提供独立信息;求稳主导说明个体差异被抹平,判别力已失。第二,本篇算出的「抵消所需漂移」(权重 0.3 需 0.040、0.6 需 0.103、0.9 需 0.094)与本领域实测漂移量级同处 0.05 到 0.20——「恰好抵消」不是巧合而是常态区,卡在门限附近的 ICC 最可能的解释不是「尺子够稳」而是「两条偏置刚好打平」。第三,出路是改设计:做分离设计,或干脆不用 ICC 作验收、改报「重复材料与新材料的成绩差」这个原始量——后者不需额外样本,只需两种材料混排。

    关于气功外气与脏腑辨证这条主线:本文不提供任何支持外气效应存在的证据。它指出的是——这条线上用来证明「测量可靠」的那把尺,原理上不足以承担这个任务。尺子换掉之前,任何基于重测一致性的分组结论只能是待验证状态。

    证据与局限

    已验证:ICC 对不同方向的系统误差敏感度极不对称(母篇实测,本篇复现)。方法学侧早有定论:PMID 11091941 用三组模拟对比 ICC、Pearson 相关、配对 t,指出三者在系统偏置下结论互相冲突、ICC 是唯一能反映偏置的,但仍须配合 F 检验确认偏置是否真存在;PMID 31329615 给出判别规则——绝对一致 ICC 与一致性 ICC 近乎相等则偏置可忽略,一致性明显高于绝对一致则不可忽略,两者必须并报。

    部分验证:两类机制确实同时存在,但各自占比从未被同时测量。记忆侧——PMID 22109344 在 24 位放射科医师、两次判读、半数换图的设计里检出记忆效应,且主要出现在低置信区;PMID 30840725 抽样 110 个观察者性能实验,67.6% 用了时间间隔(中位 14 天),52.9% 连用了什么防记忆手段都没写清。求稳侧也有实证,中心化倾向被反复记录且与严厉度漂移可分离。但没有一项研究在同一批数据上同时报出两者量级——这正是缺口。

    待验证:真实判读数据上两类强度的联合分布完全空白,本篇网格值都建立在「记忆为线性平滑、漂移为线性趋势」两条设定上;若真实记忆只在中低置信区生效(PMID 22109344 如此提示),等值线形状会变、抵消区须重算。漂移代理量在此规模下的实际精度未实测。求稳是否真能独立于记忆存在(漂移是否部分由记忆派生)未做因果分离。

    L1 网络状态:CORRESPONDENCE_NETWORK.md 第 202 行 E-060 标签前提(五脏辨证作分组标签,指向标签信度)登记为「冲突(未裁决)」,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)……但训练可改善(PMID 19538100)、量表化可行」;第 223 行补充「冲突(未裁决,且新增两处单位错配/精度区间错配)」。该主题尚未进入 L1 对应网络——全表对「记忆、求稳、中心化倾向、重测方向不对称」均无对应边,网络能登记「标签一致不一致」,登记不了「不一致是被记忆压住还是被求稳抹平」。

    相关阅读

    证据等级: C
    参考文献: PMID 11091941, PMID 31329615, PMID 30840725, PMID 22109344, PMID 30320531, PMID 31719075


    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。

    Memory pushes ICC up, stability pulls it back: when both opposite mechanisms are present, the figure near 0.5 proves nothing

    Evidence level: C

    ICC is a ratio. A ratio reports only the proportion of numerator to denominator, not their absolute magnitudes. When two opposing mechanisms act simultaneously, the same ratio can correspond to more than sixfold differences in reality.

    I. An Untested Hypothesis

    The logic of test-retest consistency is straightforward: if the same person judges the same materials twice and the results align, the ruler is stable. This relies on an assumption—if the two results do not match, it must be because the ruler is shaking.

    In reality, at least two pathways can produce different results, yet in opposite directions. The first is memory: on the second pass, seeing the same materials, the judge copies the previous conclusion; this increases consistency, pushing ICC upward. The second is conservatism: realizing they are repeating the same task, the judge fears being caught for contradictions and converges toward the middle range. This convergence brings scores closer, but it flattens differences, not locks signals—ICC may not move, yet discriminative power drops.

    The parent article proved ICC is highly sensitive to pure memory but nearly blind to pure conservatism. This piece asks the next layer: when both pathways run simultaneously, whose reading does the reported ICC represent. The answer is: neither; it is the remainder of their subtraction, and that remainder cannot reverse-engineer the two subtrahends.

    II. ICC’s Contour Lines Are Curves, Not Points

    Locking ICC to a single number reveals what lies beneath most directly. Measured: when ICC is locked at 0.18, the drift spans from 0.075 to 0.503—a difference of 6.7 times. Without memory, a drift of 0.075 suffices; with memory weight 0.6, it requires 0.150; with weight 0.9, it reaches 0.503.

    The reading is striking. A study reporting ICC = 0.18 could involve a judge with no memory and normal drift, or one with strong memory and drift five times larger. The former can be discounted; the latter must be discarded entirely—a drift of 0.503 means the third round differs from the first by half a standard deviation; they are not measured with the same ruler. Yet the report shows only “ICC = 0.18.”

    The four-fold test is more severe. Grouping by memory strength and drift magnitude yields four clusters; running 400 trials each and taking quartile intervals: low memory/low drift 0.121–0.227, low memory/high drift 0.067–0.137, high memory/low drift 0.131–0.234, high memory/high drift 0.120–0.219. All six pairwise comparisons overlap completely. At this common scale, the ICC distributions of “clean retest” and “already failed retest” groups are indistinguishable.

    III. Adding Metrics or Judges Cannot Recover the Signal

    The first instinct is “report more quantities.” This piece’s recovery metric directly measures drift, bypassing ICC’s ratio; its sensitivity indeed overwhelms: at weight 0.9, drift increases by 0.04, and its slope is 356 times that of ICC’s slope. Yet sensitivity does not equal discriminability. On an 110-cell grid, asking “given a set of observations, how many cells within tolerance can explain them”: using ICC alone, discriminability is zero; with dual metrics it reaches only 3.6%, averaging 14.3 candidate solutions per cell. Even raising ratio precision to ±4% (impossible at this scale) barely reaches 28.2%. This occurs because ICC drops slightly while drift rises slightly; these two actions are nearly locked in the low-memory region—equivalent to taking an additional projection almost collinear on an already flattened surface.

    The second instinct is “add more judges.” Measurement also fails: under strong memory, pulling drift from zero to 0.20, ICC displacement is +4 readers × 0.0068, −8 readers × 0.0119, +16 readers × 0.0011. The displacement hovers near zero, even flipping sign. Adding judges only compresses standard error (0.096 to 0.092, since ICC’s sampling error is primarily determined by sample size), but cannot move the displacement already near zero. This is not a precision problem but a structural one: two opposing terms travel the same channel; adding observations to that channel cannot separate them.

    A way exists to separate them, but it requires redesign: half the judges evaluate repeated materials, half evaluate new materials drawn from the same distribution. The memory arm uses repetition; the new-material arm does not; the difference in ICC between arms is the net memory effect, with drift cancelling across arms. Measured direction matches but signal-to-noise is poor—memory present yields arm difference 0.0522, memory absent 0.0104; achieving power of 0.30–0.37 requires 32 to 48 judges, and at 80% power, repeating 57 to 87 times. The principle holds, but practice is extremely costly.

    IV. What This Means for This Group

    All cross-subject comparisons in this group rest on one premise: grouping labels and physiological quantities are obtained from the same measurement, so test-retest consistency is the sole quality gate along this chain. Currently, that gate looks like this—it reports only one ratio, and when the two mechanisms oppose each other, this ratio is indistinguishable from “everything normal.”

    Three consequences follow. First, any study reporting “ICC meets threshold, so data are usable” lacks a statement: which mechanism dominated this number. Memory dominance means retest provided no independent information; conservatism dominance means individual differences were flattened, losing discriminative power. Second, the calculated “drift required for cancellation” (weight 0.3 needs 0.040, 0.6 needs 0.103, 0.9 needs 0.094) falls within the empirically observed drift range of 0.05–0.20 in this field—“exact cancellation” is not a coincidence but the normal zone; ICC values near the threshold are most likely explained not by “the ruler is stable” but by “two biases exactly balanced.” Third, the path forward is redesign: use a separated design, or abandon ICC as the acceptance criterion entirely and report instead “the raw difference between repeated and new materials”—the latter requires no extra samples, only interleaving the two material types.

    Regarding the main thread of qigong external qi and zang-fu differentiation: this article provides no evidence supporting the existence of external qi effects. It points out that the ruler used to prove “measurement reliability” along this line is, in principle, insufficient for the task. Until the ruler is replaced, any grouping conclusion based on test-retest consistency remains in a pending verification state.

    Evidence and Limitations

    Verified: ICC’s sensitivity to systematic errors in opposite directions is highly asymmetric (parent article measured, this piece reproduced). Methodology literature already established this: PMID 11091941 compared ICC, Pearson correlation, and paired t-tests across three simulated groups, showing conflicting conclusions under systematic bias; ICC alone reflects the bias but must be paired with F-tests to confirm whether bias truly exists. PMID 31329615 provides decision rules—if absolute consistency ICC and consistency ICC are nearly equal, bias is negligible; if consistency clearly exceeds absolute consistency, bias cannot be ignored; both must be reported.

    Partially Verified: Both mechanisms indeed coexist, yet their respective proportions have never been measured simultaneously. Memory side—PMID 22109344 detected memory effects in 24 radiologists performing two readings with half the images swapped, primarily in low-confidence regions; PMID 30840725 sampled 110 observers in a performance experiment, where 67.6% used time intervals (median 14 days), and 52.9% did not clearly state any anti-memory measures. Conservatism side also has empirical support, with centralization tendency repeatedly documented and separable from severity drift. However, no study has reported both magnitudes on the same dataset—this is the gap.

    Pending Verification: The joint distribution of both intensities in real judgment data is completely blank; all grid values here assume “memory acts as linear smoothing, drift acts as linear trend.” If real memory operates only in mid-to-low confidence regions (as PMID 22109344 suggests), the contour shape changes and the cancellation zone must be recalculated. The actual precision of drift proxies at this scale has not been measured. Whether conservatism truly exists independently of memory (i.e., whether drift is partially derived from memory) has not undergone causal separation.

    L1 Network Status: CORRESPONDENCE_NETWORK.md line 202, E-060 label premise (zang-fu differentiation as grouping labels, pointing to label reliability) registered as “Conflict (Unresolved)“, original text: “KB-292 (conflict): diagnostic consistency among TCM physicians in an open population is ‘low (low)’, authors call it ‘concerning’ (PMID 30383398) … but training can improve (PMID 19538100), scaling is feasible”; line 223 adds “Conflict (Unresolved, plus two new unit mismatches/precision interval mismatches)**. This topic has not yet entered the L1 correspondence network—the entire table lacks edges corresponding to “memory, conservatism, centralization tendency, asymmetric retest direction”; the network can register “labels consistent/inconsistent” but cannot register “inconsistency is suppressed by memory or flattened by conservatism.”

    Related Reading

    Evidence Level: C
    References: PMID 11091941, PMID 31329615, PMID 30840725, PMID 22109344, PMID 30320531, PMID 31719075


    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.

  • 洗掉记忆的那把刷子自己会掉毛:填充材料同时挪动位置和分辨率,而判定它无害的那条线比噪声还小

    Evidence level: C

    为了消除记忆而插入的填充材料,本身就是两种偏置的同时上线:一种把判断线整体挪位,一种把判读的噪声放大。它们在合计通过率上会互相抵消,而用来判定”无害”的那条阈值线,比这项设计能分辨的噪声还小。

    一、填充改的是两件事,不是一件

    填充材料与主材料难度不同,判读者在大量填充后回到主材料,会有两种后果。一种是位置挪了:主材料显得更清楚,判断线整体下移,通过率上升,这叫对比效应。另一种是分辨率降了:判得更粗,每次判读的噪声变大,这叫疲劳效应。前者改均值,后者改方差,两者不是同一件事,也不能用一个数字概括。

    这不是推测。顺序效应的方向由前后刺激的相似性决定——越相似越同化,越不相似越对比(PMID 24349226);即时语境的均值会直接改变评分水平,且符合范围-频率模型(PMID 2319444);而时间在任务上导致表现下降,恰恰在简单重复任务上最明显(PMID 23163491 的元分析)。

    二、两个机制在合计通过率上会精确抵消

    按三成材料贴阈的构型做模拟:只放对比(判断线偏移 0.20 个噪声单位),合计通过率升 3.04 个百分点;只放疲劳(噪声放大 1.3 倍),合计通过率降 1.71 个百分点;两者同时放,读到的只有 +1.21。

    这个 +1.21 容易被误读成”对比很弱”。实际上它是两个方向相反的机制各干各的之后的余数。更麻烦的是,这个余数可以恰好为零:对比 0.10 配疲劳 1.3,合计 −0.22 个百分点;对比 0.09 配疲劳 1.25,合计 −0.07 个百分点。

    于是合计通过率近零至少有三种来源:真的没偏置、单一机制被抵消、两个机制等量对冲。仅仅看合计数字,一个都分不开。

    三、界线在”贴阈材料占多少”

    同一套 40 例填充,效果不是常数。贴阈材料占一成时,疲劳让合计通过率降 2.38 个百分点,对比只升 1.77——疲劳压过对比。占比升到七成,对比升 5.29,疲劳只剩 0.22——对比主导二十四倍。

    道理很直白:对比是把整条判断线挪一点,对全部材料一视同仁;疲劳是让噪声变大,噪声只在贴阈材料上才改变结果——离阈值远的材料,噪声再大也翻不过来。

    所以”填充有害还是无害”不是材料的属性,是主材料构成的函数。同一套填充,贴阈占七成的材料上通过率虚高五个百分点,贴阈占一成的材料上虚低两个多百分点,方向都反了。这条界线可以事先算:主材料里与阈值距离在一个噪声单位内的比例,就是疲劳项的作用面;这个比例低于约两成,疲劳就压过对比。

    四、判”无害”的那条线,比噪声还小

    原方案有一格判据是”通过率升高超过 0.1 个标准差即视为以一种偏置换另一种”。若这个标准差取读者之间逐人通过率变化的离散度(模拟值 0.047),阈值就是 0.47 个百分点。而 15 位读者合并后的抽样误差是 1.20 个百分点——是阈值的 2.6 倍。

    结果是:完全不设偏置时,仍有 71.4% 的实验被判成”升高超过阈值”。要真正检出 0.1 个标准差,每臂需 768 位读者;三臂做多重比较校正后,检出 0.43 个标准差需每臂 57 位。三臂各 15 人的设计,只能检出 0.65 个标准差以上的效应。

    一个必然触发的判据不是判据,是噪声探测器。

    五、换用一致性系数这条退路也堵

    另一格判据是”重测一致性显著低于无填充臂,差超过 0.2″。实测一致性系数对记忆极敏感:记忆复用比例 0 时 0.52、0.3 时 0.67、0.6 时 0.81、0.9 时 0.95,量级足够。

    但对疲劳几乎全盲:噪声放大到 1.15 倍,系数从 0.811 变到 0.811;放大到 1.3 倍才 0.808。变化量 0.003,而单次实验的标准差是 0.025,完全淹没。原因在系数的构造:噪声放大同时抬高观察一致率与期望一致率,边际校正把两者抵掉了。

    所以”差超过 0.2″这一格两个方向都不成立:实际量级是 0.003 而不是 0.2,而所用的指标本身看不见这个机制。

    六、能站住的只有分层

    把材料按与阈值的距离预先分成贴阈、明确两档,各自算通过率变化:贴阈层对比 +6.80、疲劳 +0.62;明确层对比 +1.35、疲劳 −2.59。两层的符号组合”正、负”只在两个机制同时在场时出现,单一机制只给”正、正”或”零、负”。这是目前唯一能在 15 位读者规模上读出来的指纹。

    代价是功效:贴阈层只有 z≈5.5、明确层 z≈9.8,且真实数据里贴阈层离散度只会更大。但这是可以事先声明的分层,不是事后挖出来的。

    证据与局限

    已验证:洗脱期不足时遗留效应改变的是效应量而非假阳性率,这在交叉设计的方法学文献里有明确推导(PMID 38563531);多期设计里无洗脱期的遗留效应与相邻暴露共线,需专门模型才估得出(PMID 36762453);顺序效应方向由前后刺激相似性决定(PMID 24349226);即时语境均值改变评分水平并符合范围-频率模型(PMID 2319444、PMID 23498366);时间在任务上导致简单重复任务表现下降(PMID 23163491 元分析)。

    部分验证:填充材料在判读类任务中的具体量级——本文全部数值来自模拟,尚无同类实测。重测一致性系数的量级有先例(PMID 18560139,一周重测 0.590),但那是同材料同任务、不含填充。

    待验证:① 贴阈占比”两成”这条界线;② 疲劳的噪声放大在真实判读里是否也是 1.3 倍;③ 分层后贴阈层的功效是否真能到 z 大于 5。

    相关阅读

    证据等级: C
    参考文献: PMID 38563531, PMID 36762453, PMID 24349226, PMID 2319444, PMID 23498366, PMID 23163491, PMID 18560139
    L1 网络状态:该主题尚未进入 L1 对应网络 —— EDGE_TABLE_science-yixue.md 共 63 行、概念止于 E-061,检索”填充/洗脱/间隔设计/顺序效应/疲劳”无任何对应边;CORRESPONDENCE_NETWORK.md 第 224 行 E-060 标签前提为 conflict 状态,仅裁决标签信度与训练改善,与判读顺序无关。


    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 brush that washes away memory sheds its own bristles: the filler material shifts position and resolution, while the line that judges it harmless is smaller than noise.

    Evidence level: C

    Inserted fillers to eliminate memory inherently activate two simultaneous biases: one shifts the decision line globally, and the other amplifies reading noise. While they cancel each other out regarding total pass rate, the threshold used to judge “harmlessness” is smaller than the noise this design can resolve.

    I. Fillers Alter Two Things, Not One

    Filler materials differ in difficulty from primary materials. After extensive filling, readers return to the primary material, resulting in two consequences. First, position shifts: the primary material appears clearer, the decision line moves downward overall, and pass rate increases; this is called the contrast effect. Second, resolution decreases: judgment becomes coarser, and reading noise increases per trial; this is called the fatigue effect. The former alters the mean, while the latter alters the variance; they are not the same thing and cannot be summarized by a single number.

    This is not speculation. The direction of the sequence effect is determined by the similarity of preceding and succeeding stimuli—the more similar, the greater assimilation; the less similar, the greater contrast (PMID 24349226). The mean of the immediate context directly alters rating levels and conforms to the range-frequency model (PMID 2319444). Furthermore, time-on-task leads to performance decline, most notably in simple repetition tasks (meta-analysis in PMID 23163491).

    II. The Two Mechanisms Precisely Cancel Out in Total Pass Rate

    Simulating a configuration where 30% of materials are threshold-proximate: applying only contrast (decision line shift of 0.20 noise units) increases total pass rate by 3.04 percentage points; applying only fatigue (noise amplification by 1.3×) decreases total pass rate by 1.71 percentage points; applying both simultaneously yields only +1.21.

    This +1.21 is easily misread as “weak contrast.” Actually, it is the remainder after two opposing mechanisms operate independently. More problematically, this remainder can be exactly zero: contrast of 0.10 paired with fatigue of 1.3 yields −0.22 percentage points; contrast of 0.09 paired with fatigue of 1.25 yields −0.07 percentage points.

    Thus, near-zero total pass rate has at least three sources: no bias at all, a single mechanism being canceled, or two mechanisms offsetting equally. Looking only at the total number makes it impossible to distinguish any of these.

    III. The Threshold Lies in “Proportion of Threshold-Proximate Materials”

    With the same set of 40 filler cases, the effect is not constant. When threshold-proximate materials account for 10%, fatigue reduces total pass rate by 2.38 percentage points while contrast increases it by only 1.77—fatigue overwhelms contrast. When the proportion rises to 70%, contrast increases by 5.29 while fatigue drops to only 0.22—contrast dominates twenty-fourfold.

    The logic is straightforward: contrast shifts the entire decision line slightly, treating all materials equally; fatigue increases noise, which only changes outcomes for threshold-proximate materials—materials far from the threshold cannot flip even with high noise.

    Therefore, “is filler harmful or harmless” is not a property of the materials themselves, but a function of the primary material composition. With the same set of fillers, pass rate is artificially inflated by five percentage points on materials where 70% are threshold-proximate, but artificially deflated by over two percentage points on materials where only 10% are threshold-proximate, reversing the direction entirely. This boundary can be calculated in advance: the proportion of primary materials within one noise unit of the threshold determines the fatigue effect surface; if this proportion is below approximately 20%, fatigue overwhelms contrast.

    IV. The Line for Judging “Harmless” Is Smaller Than Noise

    The original protocol includes a criterion: “pass rate increase exceeding 0.1 standard deviations is deemed a bias substitution.” If this standard deviation refers to inter-reader variability in pass rate change (simulated value 0.047), the threshold becomes 0.47 percentage points. However, the sampling error for 15 combined readers is 1.20 percentage points—2.6 times the threshold.

    The result: without any bias, 71.4% of experiments are still judged as “pass rate increase exceeding threshold.” To truly detect 0.1 standard deviations, each arm requires 768 readers; after multiple comparison correction for three arms, detecting 0.43 standard deviations requires 57 per arm. A design with 15 readers per arm can only detect effects larger than 0.65 standard deviations.

    A criterion that inevitably triggers is not a criterion, but a noise detector.

    V. The Retreat of Using Consistency Coefficients Is Also Blocked

    Another criterion is “retest consistency significantly lower than the no-filler arm, difference exceeding 0.2.” Measured consistency coefficients are extremely sensitive to memory: at memory reuse ratio 0, ICC is 0.52; at 0.3, ICC is 0.67; at 0.6, ICC is 0.81; at 0.9, ICC is 0.95—the magnitude is sufficient.

    However, they are nearly blind to fatigue: when noise amplification reaches 1.15×, ICC drops from 0.811 to 0.811; at 1.3×, it becomes 0.808. The change is only 0.003, while the single-experiment standard deviation is 0.025—completely submerged. The reason lies in the coefficient construction: noise amplification simultaneously raises observed consistency and expected consistency, and marginal correction cancels both out.

    Therefore, “difference exceeding 0.2” fails in both directions: the actual magnitude is 0.003 rather than 0.2, and the metric itself cannot see this mechanism.

    VI. Only Stratification Can Stand

    By pre-stratifying materials into threshold-proximate and clearly distant groups based on distance from the threshold, we calculate pass rate changes separately: for the threshold-proximate layer, contrast +6.80 and fatigue +0.62; for the clear layer, contrast +1.35 and fatigue −2.59. The combination of signs “positive, negative” appears only when both mechanisms are present; a single mechanism yields only “positive, positive” or “zero, negative.” This is the only fingerprint readable with 15 readers.

    The cost is power: the threshold-proximate layer has z≈5.5, the clear layer z≈9.8, and in real data the threshold-proximate layer’s variance will be even larger. But this is pre-declared stratification, not post-hoc digging.

    Evidence and Limitations

    Verified: When washout period is insufficient, residual effects alter effect size rather than false positive rate; this has explicit derivation in crossover design methodology literature (PMID 38563531). In multi-period designs, residual effects without washout are collinear with adjacent exposures and require specialized models for estimation (PMID 36762453). Sequence effect direction is determined by stimulus similarity (PMID 24349226). Immediate context mean alters rating levels conforming to the range-frequency model (PMID 2319444, PMID 23498366). Time-on-task causes performance decline in simple repetition tasks (PMID 23163491 meta-analysis).

    Partially Verified: Specific magnitude of fillers in judgment tasks—all values here are simulated, with no comparable real measurement. Retest ICC magnitude has precedent (PMID 18560139, retest in one week ICC=0.590), but that involved same materials and tasks without fillers.

    To Be Verified: ① the “20%” boundary for threshold-proximate proportion; ② whether fatigue’s noise amplification in real judgment is indeed 1.3×; ③ whether stratified threshold-proximate layer power can truly reach z>5.

    Related Reading

    Evidence Level: C
    References: PMID 38563531, PMID 36762453, PMID 24349226, PMID 2319444, PMID 23498366, PMID 23163491, PMID 18560139
    L1 Network Status: This topic has not yet entered the corresponding L1 network — EDGE_TABLE_science-yixue.md contains 63 rows, concepts stop at E-061; searching for “filler/washout/interval design/sequence effect/fatigue” yields no corresponding edges; CORRESPONDENCE_NETWORK.md row 224, E-060 label prerequisite is conflict status, only adjudicating label credibility and training improvement, unrelated to reading order.


    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.

  • 改连续打分救不了三档:省下的量化误差本就不计入读者分歧,真正吃亏的是另一种场景

    Evidence level: C

    三档的量化误差是”共享”的,它压低了 ICC 却从不进入读者间分歧;因此换成连续打分,读数会掉,真实精度也不一定赚——账要分两个口径算,不能合成一笔。

    正文

    三档设计的读数天花板不够高,怎么办?最顺手的答案是”改连续打分”——档位是离散的,把连续量切成三格必然丢信息,而连续打分不丢。听上去是无损升级。

    但这个推理漏了一件事:量化误差和读者间分歧并不是同一笔账。 误差至少有两种来源。一种是”尺子本身刻得粗”,它对所有判读者一视同仁——大家都用同一组切点,所以这部分误差不会让两个判读者互相矛盾,只会让所有人一起偏离真值。另一种是”每人心里那把尺子不一样”——零点偏高或偏低、量程压得紧或放得松,这部分才会变成判读者之间的分歧。

    读者间一致性指标(ICC、kappa 一类)只看得见后一种。它把”两人读数差多少”当成误差的全部,而共享的那部分粗刻度,因为两人都有、且方向一致,相减时互相抵消,于是一分钱也不算进分歧里。这就是问题所在:三档的量化误差一直在替三档”藏拙”,让它的 ICC 显得比真实精度好看。 换成连续打分,这把共享的粗尺子没了,ICC 会立刻掉下来——不是因为连续打分更差,而是它不再享有那层隐身。

    这笔账算出来是多少

    把这个问题形式化:设真实状态服从标准正态,三档用等频切点分成三格,每格用格内条件期望作代表值。这样算出的共享量化误差方差是 0.206 个标准差平方(约 0.45 个标准差),明显大于均匀量化理想值”步长平方除以十二”给出的 0.062——因为三格的边缘格开口很大,条件期望对极端值的代表性差。

    现在让两位判读者各带一份方差 0.25(即 0.5 个标准差)的知觉噪声。这时三档的真实误差是 0.467,读数算出 0.793。若改用连续打分,并刻意让它达到完全相同的真实误差 0.467,它的读者间分歧就等于 0.467 全部(没有共享部分可抵消),读数只有 0.681。同一份真实精度,读数差了 0.111。 这 0.111 不是连续打分的缺陷,是三档那 0.206 从未被计数。

    反过来看更刺人:若要求连续打分也报出 0.793,它必须把真实误差压到 0.261——也就是比三档准 44%,才能打平同一个数字。所以”读数相同”绝不等于”精度相同”,两者相差一个共享量化误差的距离。

    那连续打分到底亏不亏

    这才是原题要的判决。模型里,连续打分给每位判读者的读数加上一个固定的零点偏移(标准差 sb)和一个量程缩放偏差(sg),再加知觉噪声。三者独立,真实误差就是三者平方和。设知觉噪声 0.5、量程偏差 0.1,连续打分要优于同噪声水平的三档,零点偏移必须小于 0.453 个标准差。

    这个门槛不算苛刻,但也不宽松。文献报告过的判读者严厉度差异,通常在一个到两个档位之间摆动;三档的档步长是 0.86 个标准差,换句话说,只要判读者的习惯零点差异超过半个档位,连续打分就白换了。 知觉噪声升到 0.7 时门槛掉到 0.293,几乎只剩三分之一档的余量。

    文献站哪一边

    实测证据是分裂的,而且分裂得很规整。骨折愈合的影像评分研究里,十位临床医生对 58 例(含 14 张重复片)同时用分类量表和 10 厘米视觉模拟线评分,连续打分的判读者间相关是 0.49,分类量表的加权 kappa 只有 0.25(PMID 29351044)。听觉感知的嗓音偏离度研究也同向:六位专家评 205 例,视觉模拟线的判读者间一致性高于数字分类(PMID 26222946)。但鱼类活力评分的研究结论相反:三位评分者对 304 尾鱼同时用三档分类和连续模拟线打分,连续打分没有改善活力指数与延迟死亡的拟合,作者直接建议”用三档、别用连续”(PMID 28704390)。更早的精神科症状量表研究则给出更冷的答案:26 个条目分别用分类和模拟量表评分,再把模拟结果重切成 2、3、4 档,一致性水平不受量表形式影响(PMID 515274)。

    三个方向都有,说明”连续 vs 分类”本身没有单一答案——它取决于该场景里零点偏移相对量化误差谁更大。这正是本轮算出的门槛在起作用,而不是哪一派更对。

    还有一条更硬的路

    真正值得记的不是”哪个更好”,而是操作后果:只用一致性指标来选量表形式,会系统性地偏向粗量表。 因为粗量表有一块误差不计入指标,分数天然虚高。临床能力评估的研究已反复撞到这件事——评分者严厉度差异会明显改变学生成绩与及格判定,而给评分者反馈自己的评分分布、试图校准,效果并不显著,严厉度还会随时间漂移(PMID 28587597);用多面 Rasch 模型拆解 173 名学生的临床技能评分,能直接量出严厉度差异、评分不一致、以及”只用高档位”这种量程收缩,四类错误各自独立存在(PMID 17693619)。

    所以选量表形式的正确问法不是”哪个读数高”,而是”我的判读者零点差异有多大,我的量化误差有多大,哪笔账更贵“。前一笔记进指标,后一笔不进——必须分开算,相加才是真实精度。

    证据与局限

    已验证: 在明确设定的生成模型下(三档等频切点、条件期望解码、判读者零点与量程为常数偏差),”共享量化误差不计入读者间分歧、连续打分无此豁免”是可解析验证的恒等式,不是经验猜测;仿真与解析式在 0.001 量级内吻合。文献层面,”量表形式改变一致性读数”这一现象已有多个独立场景的实测报告(PMID 29351044、26222946、28704390、515274)。

    部分验证: 零点偏移的实际量级。文献确认判读者严厉度差异普遍存在且难以校准(PMID 28587597、17693619),但没有给出可直接换算到”标准差单位”的分布参数,因此文中 0.453 这个门槛只能作为敏感性分析的界,不能当既定常数使用。

    待验证: ①本文的量化误差计算依赖”等频切点 + 条件期望解码”,实际量表若用不等频切点或简单取档序号,q2 会变,门槛随之移动;②文献里连续打分占优的三项研究(影像、嗓音)与不占优的两项(鱼类、精神科)之间,是否真能用零点偏移量级解释,还是另有第三种机制(如任务本身的可分性),目前无法从摘要层面裁决;③共享量化误差是否也影响重测信度、以及影响幅度是否与读者间口径一致,尚未计算。

    相关阅读

    证据等级: C
    参考文献: PMID 29351044, PMID 28704390, PMID 515274, PMID 28587597, PMID 17693619, PMID 26222946

    L1 网络状态:该主题尚未进入 L1 对应网络。最近邻边为 E-060 标签前提(标签信度),其状态原文为「冲突(未裁决,且新增两处单位错配/精度区间错配)」——KB-292 报中医师间诊断一致性在开放人群中「低(low)」(PMID 30383398),而同系列指出加权系数在「中低一致性 + 大样本 + 类别数多」时精度显著下降(PMID 33853549)。该边只登记「标签信度够不够用」,未登记「量表形式如何改变信度读数的记账口径」;L1 网络中「零点」「量程」「连续打分」三词命中数均为 0。


    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。

    Continuous Scoring Cannot Save Three Tiers: The Saved Quantization Error Is Already Excluded from Reader Disagreement, and the Real Loss Occurs in a Different Scenario

    Evidence level: C

    The quantization error of three tiers is “shared”; it depresses ICC but never enters reader disagreement. Therefore, switching to continuous scoring causes the reading score to drop, and true precision is not necessarily gained—the accounts must be calculated on two different scales, not merged into one.

    Body

    The ceiling of the reading score for three-tier designs is insufficient; what do we do? The most convenient answer is “switch to continuous scoring”—since tiers are discrete, chopping a continuous quantity into three bins inevitably loses information, whereas continuous scoring does not. This sounds like an upgrade without loss.

    But this reasoning misses one thing: quantization error and reader disagreement are not the same account. Error has at least two sources. One is “the ruler itself is coarsely marked,” which applies equally to all judges—everyone uses the same set of cutpoints, so this part of the error does not cause two judges to contradict each other; it merely causes everyone to deviate from the true value together. The other is “each person’s internal ruler is different”—zero points shifted high or low, ranges compressed tight or stretched loose; this part becomes the disagreement between judges.

    Reader consistency indices (ICC, kappa, etc.) only see the latter. They treat “how much do two readings differ” as the entirety of error, while the shared part of coarse scaling, since both judges have it and it points in the same direction, cancels out when subtracted, so not a cent is counted into disagreement. This is the crux of the problem: the quantization error of three tiers has been hiding the shortcomings of three tiers, making its ICC appear better than its true precision. Switching to continuous scoring removes this shared coarse ruler, so ICC immediately drops—not because continuous scoring is worse, but because it no longer enjoys that layer of invisibility.

    How much does this account come to?

    Formalizing the problem: suppose the true state follows a standard normal distribution, three tiers use equal-frequency cutpoints to divide into three bins, and each bin uses the conditional expectation within the bin as its representative value. The resulting shared quantization error variance is 0.206 standard deviation squared (approximately 0.45 standard deviations), clearly larger than the uniform quantization ideal value of “step squared divided by twelve” yielding 0.062—because the edge bins of three tiers have large openings, and conditional expectations poorly represent extreme values.

    Now suppose two judges each carry perceptual noise with variance 0.25 (i.e., 0.5 standard deviations). Then the true error of three tiers is 0.467, and the reading score calculates to 0.793. If we instead use continuous scoring and deliberately make it achieve the exact same true error of 0.467, its reader disagreement equals the full 0.467 (no shared part to cancel), so the reading score is only 0.681. The same true precision yields a reading difference of 0.111. This 0.111 is not a defect of continuous scoring; it is because the 0.206 from three tiers was never counted.

    Looking at it the other way is even more striking: if we demand continuous scoring also report 0.793, it must suppress the true error to 0.261—44% more accurate than three tiers—to match the same number. Thus “same reading” never means “same precision”; they differ by the distance of shared quantization error.

    So does continuous scoring actually lose out?

    This is the judgment the original question seeks. In the model, continuous scoring adds to each judge’s reading a fixed zero-point offset (standard deviation sb) and a range scaling bias (sg), plus perceptual noise. These three are independent; true error is the sum of their squares. Suppose perceptual noise is 0.5 and range bias is 0.1; for continuous scoring to outperform three tiers at the same noise level, the zero-point offset must be less than 0.453 standard deviations.

    This threshold is not overly strict, but it is not loose either. Literature reports of judge severity differences typically swing between one and two tiers; the step size of three tiers is 0.86 standard deviations, meaning if the judges’ habitual zero-point difference exceeds half a tier, switching to continuous scoring is pointless. When perceptual noise rises to 0.7, the threshold drops to 0.293, barely leaving room for one-third of a tier.

    Where does the literature stand?

    Empirical evidence is split, and split in a regular way. In fracture healing imaging scoring studies, ten clinicians simultaneously used categorical scales and 10 cm visual analog scales on 58 cases (including 14 repeated images); the correlation between continuous scorers was 0.49, while the weighted kappa of categorical scales was only 0.25 (PMID 29351044). Auditory perception studies of voice deviation also align: six experts rated 205 cases, and consistency among visual analog scorers exceeded that of digital categorizers (PMID 26222946). But fish vitality scoring studies concluded the opposite: three raters scored 304 fish simultaneously using three-tier categorization and continuous analog lines, and continuous scoring did not improve the fit of vitality index to delayed mortality; authors directly advised “use three tiers, don’t use continuous” (PMID 28704390). An earlier psychiatry symptom scale study gave an even colder answer: 26 items were scored using both categorical and analog scales, then the analog results were re-chopped into 2, 3, or 4 tiers, and consistency levels were unaffected by scale form (PMID 515274).

    All three directions exist, showing that “continuous versus categorical” itself has no single answer—it depends on whether the zero-point offset relative to quantization error is larger in that scenario. This is precisely where the threshold calculated above comes into play, not which side is more correct.

    There is still a harder path

    What truly matters to remember is not “which is better,” but the operational consequence: using consistency indices alone to choose scale form systematically favors coarse scales. Because coarse scales have a portion of error not counted in the index, their scores are naturally inflated. Clinical skills assessment studies have repeatedly bumped into this—rater severity differences significantly alter student scores and pass/fail judgments, while giving raters feedback on their own score distributions to attempt calibration has not significantly improved results; severity still drifts over time (PMID 28587597); decomposing clinical skills scores of 173 students using the many-facet Rasch model directly measures severity differences, rater inconsistency, and “using only high tiers” range contraction; these four errors exist independently (PMID 17693619).

    Thus the correct question for choosing scale form is not “which reading is higher,” but “how large are my raters’ zero-point differences, how large is my quantization error, which account costs more.” The first enters the index; the second does not—must be calculated separately, added together for true precision.

    Evidence and Limitations

    Verified: Under clearly specified generative models (three tiers with equal-frequency cutpoints, conditional expectation decoding, raters’ zero points and ranges as constant biases), “shared quantization error does not enter reader disagreement, continuous scoring lacks this exemption” is an analytically verifiable identity, not empirical guesswork; simulations and analytical formulas agree within 0.001 magnitude. At the literature level, “scale form changes consistency readings” has been reported in multiple independent scenarios (PMID 29351044, 26222946, 28704390, 515274).

    Partially Verified: The actual magnitude of zero-point offsets. Literature confirms rater severity differences are ubiquitous and hard to calibrate (PMID 28587597, 17693619), but does not provide distribution parameters directly convertible to “standard deviation units”; therefore the threshold of 0.453 in this text can only serve as a sensitivity analysis bound, not a fixed constant.

    To Be Verified: ① This paper’s quantization error calculation relies on “equal-frequency cutpoints + conditional expectation decoding”; if actual scales use unequal-frequency cutpoints or simple bin indices, q2 changes and the threshold shifts accordingly; ② among the three studies where continuous scoring prevailed (imaging, voice) and two where it did not (fish, psychiatry), whether zero-point offset magnitude alone can explain the difference, or whether a third mechanism (such as task divisibility itself) is at play, cannot be adjudicated from abstracts alone; ③ whether shared quantization error also affects test-retest reliability, and if so, whether the effect magnitude matches the reader-disagreement scale, has not been computed.

    Further Reading

    Evidence Level: C
    References: PMID 29351044, PMID 28704390, PMID 515274, PMID 28587597, PMID 17693619, PMID 26222946

    L1 Network Status: This topic has not yet entered the L1 corresponding network. The nearest edge is the E-060 label premise (label credibility), whose original status reads “Conflict (unresolved, and two additional unit mismatches/precision interval mismatches added)“—KB-292 reports that diagnostic consistency among TCM practitioners in open populations is “low” (PMID 30383398), while the same series indicates weighted coefficients suffer significant precision decline under “medium-low consistency + large sample + many categories” (PMID 33853549). This edge only registers “whether label credibility suffices,” not “how scale form changes the accounting scale of credibility readings”; in the L1 network, hit counts for “zero point,” “range,” and “continuous scoring” are all zero.


    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.

  • 那把秤量的根本不是「稳不稳」:它只看差值抖得有多散,而两个人一起偏,它一点都看不见

    Evidence level: C

    上一篇把「δ 稳不稳」的判据画成「批次间标准差 ÷ 抽样地板」,并指出地板不在 1。这一篇要问的是更前一步:这个比值里的分子是「差值的标准差」,而标准差对「两个人一起偏」完全无感。所以它不是「稳不稳」的秤,只是一台「抖不抖」的尺。

    正文

    上一篇(#1410)把「δ 买不买得回来」压成了一个可算的比值:批次间 δ 的标准差 ÷ 含 ρ 的二项地板。地板是 √(2−2ρ) 而非 1,ρ=0 时 1.414,判据线画在 1.5,离地板只剩 6%。

    这件事成立。但往下走会遇到一个从没被问过的问题:比值大于 1.5,到底证明了什么?

    答案比想象的窄:它只证明「δ 在各批次之间抖得比抽样噪声大」。而「抖得大」和「偏得稳」是两件完全不同的事。

    一、标准差对「一起偏」是瞎的

    取两位读者的主档比例 p_A 与 p_B,δ = p_A − p_B。这里有个初等但致命的性质:δ 的标准差只与 p_A、p_B 的散布有关,与它们的共同水平无关。

    若把 p_A 与 p_B 同时抬高同一个常数 c,δ 逐批不变,标准差一分不动。同理,若两位读者各带一个固定风格偏移 b_A、b_B——这正是「稳定成分」最直白的形态——δ 每批都是同一个常数,标准差恰好是零。

    仿真直接照见这一点。在同一潜变量真值上,让两位读者各带固定门槛偏移(批批不变,这就是最纯粹的「稳定风格」),跑十二批次、每批一百二十例:δ 的批间标准差与完全没有偏移时几乎一模一样,差距在小数点后第二位。二十个百分点的真实系统偏,在分子里一个数字都没挪动。

    不是「偏小」,是零敏感度。这也解释了 #1410 那张表里 τ 那一列的量级为何总小得可疑:它测的从来不是稳定偏移的大小,而是批次漂移的散度。

    二、它真正在测的是「漂」不是「偏」

    把两种形态并排跑,差别立刻摊开(sg=0.5、n=120、B=12,纯噪声基线 S=4.62pp、越 1.5 的概率 1.3%):

    真实幅度 形态 S(pp) 越 1.5 的概率 符号一致率≥0.75
    10 pp 固定风格(不漂) 4.78 1.3% 53.0%
    10 pp 批次漂移 6.51 30.2% 9.8%
    20 pp 固定风格(不漂) 4.87 0.5% 68.5%
    20 pp 批次漂移 9.83 84.3% 11.7%

    读到的是两条完全相反的曲线。二十个百分点的固定偏差,越线率只有 0.5%——比纯噪声的 1.3% 还低,因为恒定偏移把方差挤小了。同样的二十个百分点,若换成每批独立漂移,越线率飙到 84.3%。

    所以这个比值是一台漂移探测器,而不是一台偏差探测器。 它对「两位读者一直差一格」这种最该被警惕的情形,灵敏度是零。

    三、零分布的第 37 类缺口:差值的散布 vs 差值的位置

    问题还不在灵敏度低,而在它的零分布与它想回答的问题根本不是同一个分布。

    「δ 稳不稳」问的是:δ 的期望值是否偏离零、且偏离得足够一致。这是一个关于位置的问题,零假设是「δ 的均值等于 0」。而比值统计量用的是 δ 的散布,零假设是「δ 的方差等于抽样方差」。两个假设在数学上互相独立:δ 可以均值极大而方差极小(完全稳定的偏差,本判据判「可补偿」),也可以均值恰好为零而方差极大(纯随机漂移,本判据判「有稳定成分」)。

    这是判据设计里最易被忽略的一类缺口——不是「估得准不准」,而是统计量瞄准的靶心与要回答的问题不是同一个。

    补它不需新数据,只需换统计量:看 δ 的符号在批次间是否一致。同样十二批次,纯噪声下「符号一致率 ≥ 0.75」的误判率 9.5%;十个百分点固定偏差把它推到 53.0%,二十个百分点到 68.5%;而十个百分点批次漂移只有 9.8%、二十个百分点 11.7%。一升一不升,形态就分开了。

    四、两个统计量必须并报

    单独换成符号一致率也不够。它是离散的、样本效率低:十二批次时,二十个百分点固定偏差的命中率也只有 68.5%,还有三成漏掉。而比值统计量在漂移场景下确实更强(32% vs 12%)。

    结论不是二选一,而是两者测的是不同的事,缺一就读不全:

    • 比值大、符号乱 → 批次漂移主导,δ 每批往不同方向拉,期望抵消但每批都在锯天花板;
    • 比值小、符号齐 → 固定风格偏移主导,δ 是个几乎不变的常数,这才对应 #1396 里「买不回来」的那一半;
    • 比值大、符号齐 → 两者叠加,最坏情形,δ 既偏又抖;
    • 比值小、符号乱 → 真噪声,可补偿。

    只看比值,第二格会被误判成第一格之外的第一类——这正是「二十个百分点稳定偏移被判为可补偿」的现实版本。

    五、这对「五脏标签」意味着什么

    回到本组最关心的场景。用五脏辨证做分组标签时,两位医师之间的分歧有两种来源:一是诊断标准本身模糊,导致每批病人判得不一样(漂移);二是两位医师对「脾虚」的门槛本来就不在同一位置,且这个位置长期不变(固定风格)。

    前者可以靠加批次、加样本平均掉,后者加多少批次都不动——正是 #1396 说的「买不回来」。而 #1410 的比值统计量只看得见前者。若某研究报了「比值 1.2,δ 可补偿」就收工,那个更顽固的固定偏移从头到尾没被称过。

    文献侧处境相同。判读者漂移在观察评分领域被反复测到且量级惊人——四百五十八位教师的两年数据显示,漂移在最初几天极大并持续近两年,评分者非但没收敛到共同严厉度,彼此差异反而扩大(PMID 29795823)。而「固定偏移与批次漂移能否分开估」,多数 G 研究给出的答案悲观:设计上 occasion 常与 lesson 完全混淆,只能估交互项(PMID 32655177);临床评分中 rater 方差占 59%、occasion 仅 18%,但那是嵌套分解,不是独立识别(PMID 26925540)。

    证据与局限

    已验证(纯解析,可复算):δ 的批间标准差对两位读者的恒定共同偏移严格不敏感——这是标准差的定义直接给出的,不是经验发现。仿真数值(4.62 / 4.68 pp)与定义一致。

    部分验证(仿真 + 文献方向一致):比值统计量在「固定偏差」与「批次漂移」两种形态下的判别力严重不对称(20pp 时 0.5% vs 84.3%),符号一致率则反向(68.5% vs 11.7%)。文献侧只有间接印证——漂移确实存在且量级大(PMID 29795823),rater 与 occasion 在常规设计中无法独立识别(PMID 32655177、PMID 26925540、PMID 25683899),但没有任何一篇直接比较过这两种统计量的判别力。

    待验证:① 符号一致率在真实判读数据上的检验效能(本文只跑了潜变量仿真,未用实测数据);② 固定偏移与批次漂移在 8 位读者 × 12 批次的设计下能否被联合识别;③ 若固定偏移的绝对量级普遍大于漂移,则整条「加批次可补偿」的设计建议需要重估。

    明确未做的事:本文没有推翻 #1410 的地板公式(√(2−2ρ) 的推导正确),也没有声称比值统计量无用——它是漂移探测器,只是被当成了偏差探测器用。原题预设的「ρ 与 τ 同源会把比值双向推向 1」,在仿真里只得到部分支持:阈值层漂移下 corr(ρ,τ) 呈强负相关(−0.97),概率层漂移下呈 −0.99,但这是「读者越精准、同一门槛漂移打出更大边际漂移」的放大率效应,与「同一个风格离散度的两种投影」不是一回事。

    L1 网络状态:CORRESPONDENCE_NETWORK.md 全文 1274 行逐词检索,「kappa」0 次、「方差」0 次、「边际」0 次、「判读」0 次、「读者」0 次;「一致性」6 次、「信度」8 次全部落在 E-060 标签前提边,且该边状态为 conflict(未裁决)——KB-292(PMID 30383398)判中医师间诊断一致性「低」、作者称「令人担忧」,而 KB-317/334 指出该判断存在样本量单位错配(n≥50 指评定者 vs 120 指病例)与精度区间错配(「一致性 low」正落在加权系数精度最差区间)。EDGE_TABLE 概念最高止于生理概念,无一致性度量节点。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 29795823, PMID 25683899, PMID 34111197, PMID 11591433, PMID 16164474


    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 scale doesn’t measure stability at all; it only sees how scattered the difference is, so if both people drift together, it remains completely invisible.

    Evidence level: C

    In the previous post, the criterion for “δ stability” was rendered as the ratio of between-batch standard deviation to sampling floor, noting that this floor is not 1. This post asks the step prior: the numerator of this ratio is the “standard deviation of differences,” and standard deviation is completely insensitive to “both raters drifting together.” Thus, it is not a scale for “stability,” but merely a ruler for “jitter.”

    Main Text

    The previous post (#1410) compressed the question of “δ buyability” into a calculable ratio: between-batch δ standard deviation ÷ binomial floor containing ρ. The floor is √(2−2ρ) rather than 1; at ρ=0 it equals 1.414, the decision line is drawn at 1.5, leaving only 6% above the floor.

    This holds. But proceeding further reveals an unanswered question: what does a ratio greater than 1.5 actually prove?

    The answer is narrower than imagined: it proves only that “δ jitters between batches more than sampling noise.” “Jittering greatly” and “drifting steadily” are two entirely different matters.

    I. Standard Deviation is Blind to “Shared Drift”

    Take two readers’ main proportions p_A and p_B, with δ = p_A − p_B. Here lies an elementary yet fatal property: the standard deviation of δ depends only on the spread of p_A and p_B, not on their common level.

    If both p_A and p_B are raised by the same constant c, δ remains unchanged batch-by-batch, and the standard deviation does not move. Similarly, if each reader carries a fixed style offset b_A, b_B—this is the most direct form of “stable component”—δ is a constant each batch, and the standard deviation is exactly zero.

    Simulation confirms this. Given the same latent true values, let each reader carry a fixed threshold offset (unchanging across batches, representing pure “stable style”). Running twelve batches of 120 cases each: the between-batch standard deviation of δ is almost identical to the case with no offset, differing only in the second decimal place. Twenty percent of true systematic bias moves not a single digit in the numerator.

    It is not “less sensitive,” but zero sensitivity. This explains why the τ column in #1410’s table always appears suspiciously small: it never measures the magnitude of stable offset, but rather the divergence of batch drift.

    II. It Actually Measures “Drift,” Not “Bias”

    Running both forms side-by-side immediately reveals the difference (sg=0.5, n=120, B=12, pure noise baseline S=4.62pp, probability of exceeding 1.5 is 1.3%):

    True Magnitude Form S (pp) Prob > 1.5 Symbol Agreement ≥0.75
    10 pp Fixed style (no drift) 4.78 1.3% 53.0%
    10 pp Batch drift 6.51 30.2% 9.8%
    20 pp Fixed style (no drift) 4.87 0.5% 68.5%
    20 pp Batch drift 9.83 84.3% 11.7%

    The results trace two completely opposite curves. Twenty percent of fixed bias yields an exceedance rate of only 0.5%—even lower than pure noise at 1.3%, because constant offset compresses the variance. The same twenty percent, if converted to independent batch drift, pushes the exceedance rate up to 84.3%.

    Thus this ratio is a drift detector, not a bias detector. Its sensitivity to “two readers consistently differing by one unit”—the scenario most worthy of vigilance—is zero.

    III. The 37th Gap in the Zero Distribution: Spread of Differences vs. Position of Differences

    The problem is not merely low sensitivity, but that its zero distribution is fundamentally different from the distribution it aims to answer.

    “Is δ stable?” asks whether the expected value of δ deviates from zero and does so consistently. This is a question about position; the null hypothesis is “the mean of δ equals zero.” Yet the ratio statistic uses the spread of δ, with the null hypothesis being “the variance of δ equals the sampling variance.” These two hypotheses are mathematically independent: δ can have a large mean and tiny variance (completely stable bias, which this criterion would deem “compensatable”), or it can have a mean exactly zero and large variance (pure random drift, which this criterion would deem “having stable components”).

    This is the most overlooked gap in criterion design—not “how accurately it estimates,” but that the target of the statistic does not match the question to be answered.

    Fixing it requires no new data, only changing the statistic: examine whether the sign of δ remains consistent across batches. With twelve batches, under pure noise, the misclassification rate for “symbol agreement ≥ 0.75” is 9.5%; ten percent fixed bias pushes it to 53.0%, twenty percent to 68.5%; while ten percent batch drift yields only 9.8%, twenty percent 11.7%. One rises, the other does not; the forms separate clearly.

    IV. Both Statistics Must Be Reported

    Switching solely to symbol agreement is insufficient. It is discrete and sample-inefficient: with twelve batches, twenty percent of fixed bias achieves a hit rate of only 68.5%, missing thirty percent. Meanwhile, the ratio statistic is indeed stronger under drift scenarios (32% vs 12%).

    The conclusion is not a choice between two, but that they measure different things, and lacking one leaves the picture incomplete:

    • Ratio high, symbols scattered → batch drift dominates; δ is pulled in different directions each batch, expectations cancel but each batch saws the ceiling;
    • Ratio low, symbols aligned → fixed style offset dominates; δ is nearly constant, corresponding to the “unbuyable” half mentioned in #1396;
    • Ratio high, symbols aligned → both combined, worst-case scenario; δ is both biased and jittery;
    • Ratio low, symbols scattered → true noise, compensatable.

    Looking only at the ratio misclassifies the second cell as the first type outside the second—precisely the real-world version of “twenty percent stable offset deemed compensatable.”

    V. What This Means for the “Five-Organ Labels”

    Returning to the most relevant scenario in this group. When grouping labels via five-organ differentiation, discrepancies between two physicians arise from two sources: either the diagnostic criteria themselves are vague, causing inconsistent patient classification across batches (drift); or the two physicians have fundamentally different thresholds for “spleen deficiency” that remain unchanged over time (fixed style).

    The former can be averaged out by adding batches and samples; the latter remains unchanged regardless of how many batches are added—exactly what #1396 calls “unbuyable.” Meanwhile, the ratio statistic in #1410 only detects the former. If a study reports “ratio 1.2, δ compensatable” and stops there, the more stubborn fixed offset goes entirely unmentioned.

    The literature faces the same situation. Drift in raters has been repeatedly observed and found to be of remarkable magnitude in the observation-rating domain—data from 458 teachers over two years shows drift is extreme initially and persists for nearly two years, with raters not only failing to converge on common severity but their differences actually expand (PMID 29795823). Whether “fixed offset and batch drift can be separately estimated,” most G studies give a pessimistic answer: by design, occasion often completely confounds with lesson, allowing only interaction terms to be estimated (PMID 32655177); in clinical ratings, rater variance accounts for 59%, occasion only 18%, but that is nested decomposition, not independent identification (PMID 26925540).

    Evidence and Limitations

    Verified (purely analytical, reproducible): the between-batch standard deviation of δ is strictly insensitive to constant shared offset between two readers—this follows directly from the definition of standard deviation, not empirical discovery. Simulation values (4.62 / 4.68 pp) align with the definition.

    Partially Verified (simulation + literature direction consistent): the discriminative power of the ratio statistic is severely asymmetric between “fixed bias” and “batch drift” forms (0.5% vs 84.3% at 20pp), while symbol agreement reverses (68.5% vs 11.7%). The literature offers only indirect corroboration—drift indeed exists and is of large magnitude (PMID 29795823), and rater and occasion cannot be independently identified in conventional designs (PMID 32655177, PMID 26925540, PMID 25683899), but no study has directly compared the discriminative power of these two statistics.

    Pending Verification: ① the statistical power of symbol agreement on real judgment data (this paper ran only latent variable simulations, not actual measured data); ② whether fixed offset and batch drift can be jointly identified under an 8 readers × 12 batches design; ③ if the absolute magnitude of fixed offset is generally larger than drift, the design recommendation of “adding batches compensates” needs reevaluation.

    Explicitly Not Done: This paper does not overturn the floor formula in #1410 (the derivation of √(2−2ρ) is correct), nor does it claim the ratio statistic is useless—it is a drift detector, merely misused as a bias detector. The original premise that “ρ and τ share the same source, pushing the ratio bidirectionally toward 1” receives only partial support in simulation: under threshold-layer drift, corr(ρ,τ) shows strong negative correlation (−0.97), under probability-layer drift −0.99, but this is an amplification effect of “the more precise the reader, the larger the marginal drift from the same threshold,” not the same as “two projections of the same style dispersion.”

    L1 Network Status: Full text search of CORRESPONDENCE_NETWORK.md (1274 lines) yields zero occurrences of “kappa,” “variance,” “marginal,” “judgment,” or “rater”; “consistency” (6 times) and “reliability” (8 times) all fall under the E-060 label premise edge, which is in state conflict (unresolved)—KB-292 (PMID 30383398) judges inter-physician diagnostic consistency as “low,” with authors calling it “concerning,” while KB-317/334 indicates this judgment suffers from sample unit mismatch (n≥50 refers to raters vs 120 cases) and precision interval mismatch (“consistency low” falls precisely in the worst-precision weighted coefficient region). The EDGE_TABLE concept extends only to physiological concepts, with no consistency measurement node. This topic has not yet entered the L1 corresponding network.

    Related Reading

    Evidence Level: C
    References: PMID 29795823, PMID 25683899, PMID 34111197, PMID 11591433, PMID 16164474


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

  • 三档那个漂亮的 1+2E,是「只剩一个方向」造成的巧合:量表每加一档,安全区就要乘上一个小于一的数,乘到第五档基本没了

    Evidence level: C

    上一轮给出的 1+2E 之所以干净,是因为三档量表把自由度恰好压成了一维。四档起方向变成两个、三个,而真实约束是「预算总量」,不是「每个方向各自放宽」——于是每加一档,安全区不是宽一点,而是乘上一个小于一的数。

    正文

    先说结论会推翻什么。

    前一轮把判据的放宽量算成一个干净的 1+2E,E 是两位读者「一起往哪边偏」。当时有个自然的期待:四档、五档时把 E 逐维套一遍,(1+2E) 各自乘开就行。这个期待是错的,而且错得很彻底——不是差个百分之十几,是量级错。

    为什么三档会让人误以为它干净。

    三档量表里,主档之外的次档只有两个:第二档、第三档。两位读者的质量守恒各给一条约束,于是次档差向量只剩一维自由度——一个数 Δ 就描述完了。一维的东西,「允许范围有多长」和「允许范围有多大」是同一件事,所以长度比 1+2E 同时也是体积比。干净不是因为它对,是因为没有第二个方向来暴露它。

    四档时次档有三个,自由度为二;五档自由度为三。判据也变了形状:不再是「一个数的绝对值不超过 δ/2」这条线,而是「所有次档差同号且预算够」,这是若干个不等式围出来的一块区域。

    算出来的东西比预想的简单。

    把两位读者的次档分配看作在单纯形上随机摊开,问「恰好不越线」的概率。设主档占比差为 δ,较大的一侧主档占比为 1−s_B。扫 k=3 到 6、多组 (δ, 主档均值),实测出一个精确到四位小数的律:

    每一档的收缩因子恒为 δ 除以 s_B,且第 k 档的概率正好是这个因子的 k−2 次方。八组参数、每组两百万次抽样,实测与闭式在第三四位上吻合(δ=0.05、主档均值 0.88 时,因子 0.0552,三档 0.0552、四档 0.0031、五档 0.0002 不到、六档已经落在万分之一以下)。

    这个因子不含 E。 E 在多档里连影子都不剩:把 E 按五分位分箱,看每一箱内的安全率,四档 0.1601/0.1601/0.1601/0.1601/0.1603——五个箱子差不到万分之二。也就是说,「两人一起往哪边偏」这条轴在三档里能决定放宽多少,在四档里对总体安全份额毫无分辨力。它没有消失,只是被 k−1 个方向平均掉了。

    三种搬运三档结论的做法,各自漏多少。

    折档法(把 k 档并成「第一档/其余」当三档读):四档、主档差比 0.5 时,折档放行的样本里有 66.7% 其实越了线;五档升到 88.1%;六档 95.5%。档数越多、折得越狠、错得越离谱——因为合并把「某一次档单独超标」抹平成了「合计不超」。

    逐维朴素法(要求每个次档差绝对值都不超过 δ):四档 65.3%、五档 88.7%、六档 96.2% 的放行是假的。根因相同:k 条差各自不超 δ,加起来可以到 kδ,而预算只有 δ。

    逐维套 (1+2E):E=0.20、六档时给出 5.38 倍放大,而真实放大是 1 倍。方向完全反了——它把判据放松了近五倍,实际该做的是收紧。

    成本落在样本量上,而且是爆炸式的。

    既然安全区的份额按那个因子的 (k−3) 次方缩,样本量就得按倒数放大。δ=0.05、主档均值 0.90 这组很现实的高一致参数下:三档设计 n=100,四档要 1850,五档要 34225,六档要 63 万。δ=0.05、主档均值 0.80 时四档 1650、五档 27225。「多加一档让判定更细」在统计上不是免费的加细,是换了一个量级的代价。

    唯一没被推翻的是那条纯观测判据。

    上一轮末尾给的替身判据——TVD = δ 等价于「每一条次档差都不为正」——这一轮在 k=3/4/5/6、多种主档差下各四十万组随机边际上全部核对通过,一致性 100.00%,通过率逐位相同。它不需要估 E,不需要知道谁是 A 谁是 B,不需要逐维套倍数。

    配套要报的只有两个数:主档差 δ,和较大主档那一侧的剩余质量 s_B。后者决定因子,前者已经在报告里。

    这件事撞上了研究内核里那条还没裁的冲突边。

    我们网络有一个分组前提:用五脏辨证给受试者贴标签。这条前提现在挂在 conflict 状态,双方都拿得出证据。悲观一侧是中医医师间的诊断一致性在开放人群中「低」,原作者用了「令人担忧」这个词(PMID 30383398)。乐观一侧来自同一系列:那个「低」有相当部分是统计量误用造成的,换 Gwet AC2 重算后线性加权一致率 0.80、偶然校正后落在 0.67 至 0.73,且训练可改善(PMID 19538100,PMID 29851495 专门论证 AC2 才是自由边际数据该用的统计量)。两边说的是同一批数据,分歧在用了哪个系数。 而本篇说的正是同一类事:一个判据的松紧,可能整个来自它假设了几维,而不是数据本身。所以那句「一致性低所以标签不可靠」在换档数、换系数之后还能不能站住,目前不能当已裁决的结论用。

    证据与局限

    已验证(纯数学,非实证):k 档下「恰好不越线」的概率按 (δ/s_B)^(k−2) 收缩,闭式与蒙特卡洛在八组参数、每组两百万次抽样下逐位吻合;TVD ≤ δ 与「所有次档差 ≤ 0」在 k=3 至 6 全域 100% 等价;三档的 1+2E 恰好等于一维情形的长度比,与体积比同值属退化巧合。

    部分验证:折档法、逐维朴素法、逐维套 1+2E 三种错法的误判率均在本次模拟的参数网格内算出,网格覆盖主档差比 0.3 至 1.5、档数 3 至 6;网格外的极低一致率区未扫。

    待验证:真实裁决研究里四档以上量表的主档占比与主档差实际落在哪个区,本次全部数值基于均匀随机分配的假设,真实读者有系统性偏好,分布不是均匀的,会让因子偏离;「档数越多样本量越爆炸」这条在已发表的多档一致性研究里从未被当作设计约束写出来,需回查是否有研究因此实际欠功效。

    L1 网络状态:L1 对应网络中未找到该主题(有序量表档数对一致性判据安全区体积的影响)的边。边表概念共 61 条(E-001 至 E-061),无一条登记「量表档位数」作为变量;最接近的 E-060「五脏辨证作分组标签|前提|标签信度」仅以 conflict 状态登记「标签信度够不够」(KB-292/KB-317/KB-333/KB-334),未登记「一致性系数本身的判据会随档数改变形状」。故本篇证据等级封顶 C。

    相关阅读

    证据等级: C
    参考文献: PMID 8834562, PMID 33853549, PMID 10769936, PMID 30869956, PMID 36050908, PMID 29851495, PMID 32020851

    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。

    That beautiful 1+2E on three tiers is a coincidence caused by having only one direction: each additional tier multiplies the safety zone by a number less than one, so by the fif…

    Evidence level: C

    The clean 1+2E derived in the previous round is so tidy because the three-tier scale compresses degrees of freedom into exactly one dimension. Starting with four tiers, directionality becomes two-dimensional, then three-dimensional, while the true constraint remains the “total budget,” not “each direction relaxing independently”—thus, with each added tier, the safety zone does not widen; it is multiplied by a factor less than one.

    Main Body

    First, what conclusions will be overturned.

    The previous round calculated the relaxation amount of the criterion as a clean 1+2E, where E represents the two readers’ “joint offset direction.” A natural expectation arose: apply this 1+2E dimension-by-dimension for four-tier and five-tier scales, simply multiplying out (1+2E). This expectation is wrong, and utterly so—not off by a fraction of ten percent, but off by orders of magnitude.

    Why three tiers create the illusion of cleanliness.

    In the three-tier scale, there are only two sub-tiers beyond the main tier: second tier and third tier. Mass conservation from each reader provides one constraint, leaving the sub-tier difference vector with one degree of freedom—fully described by a single number Δ. For one-dimensional quantities, “how long the allowed range is” and “how large the allowed range is” are identical; thus, the length ratio 1+2E equals the volume ratio. Cleanliness does not stem from correctness, but from the absence of a second direction to expose this degeneracy.

    With four tiers, sub-tiers become three, yielding two degrees of freedom; with five tiers, three degrees of freedom. The criterion’s shape changes: no longer “the absolute value of one sub-tier difference not exceeding δ/2,” but rather “all sub-tier differences sharing the same sign and the budget sufficing”—a region bounded by multiple inequalities.

    The calculated results are simpler than expected.

    Treating the allocation of sub-tiers for two readers as random spreading over a simplex, we ask: what is the probability of “exactly not crossing the line”? Let δ denote the main-tier proportion difference, and 1−s_B the larger-side main-tier proportion. Sweeping k=3 to 6 across multiple (δ, mean main-tier proportion) configurations yields an exact law accurate to four decimal places:

    The contraction factor for each tier is constant at δ/s_B, and the probability for tier k equals this factor raised to the power (k−2). Across eight parameter sets, each with two million Monte Carlo samples, the empirical results match the closed-form expression precisely in the third and fourth decimal places (e.g., δ=0.05, mean main-tier proportion 0.88 → contraction factor 0.0552; three tiers: 0.0552, four tiers: 0.0031, five tiers: 0.0002, six tiers already below one in ten thousand).

    This factor contains no E. E vanishes entirely across multiple tiers: binning E into quintiles and examining safety rates within each bin yields four-tier values of 0.1601/0.1601/0.1601/0.1601/0.1603—differing by less than two ten-thousandths across five bins. Thus, “the axis of joint offset direction” determines relaxation in three tiers but exerts no discriminatory power on overall safety share in four tiers. It does not disappear; it is averaged out by k−1 directions.

    Three methods of transporting the three-tier conclusion each miss something.

    • Tier-folding method (merging k tiers into “first tier / rest” as three-tier readout): At four tiers with main-tier proportion difference ratio 0.5, 66.7% of folded-allowed samples actually cross the line; five tiers: 88.1%; six tiers: 95.5%. The more tiers, the harsher the fold, the worse the error—because folding erases “one sub-tier alone exceeding” into “sum not exceeding.”
    • Naive dimension-by-dimension method (requiring each sub-tier difference absolute value ≤ δ): Four tiers: 65.3% falsely allowed; five tiers: 88.7%; six tiers: 96.2%. Root cause identical: k differences each ≤ δ sum to kδ, while budget is only δ.
    • Dimension-by-dimension applying (1+2E): With E=0.20 at six tiers, this yields a 5.38× amplification, whereas the true amplification is only 1×. The direction is completely reversed—it relaxes the criterion nearly fivefold when tightening is required.

    Cost falls on sample size, and explosively.

    Since safety zone share contracts by the (k−3) power of that factor, sample size must expand by its reciprocal. Under δ=0.05 and mean main-tier proportion 0.90—a realistic high-consistency parameter set: three tiers n=100, four tiers n=1850, five tiers n=34,225, six tiers n=630,000. At δ=0.05 and mean main-tier proportion 0.80: four tiers n=1,650, five tiers n=27,225. “Adding one tier to refine judgment” is not free refinement statistically; it incurs an order-of-magnitude cost.

    The only conclusion not overturned is the pure observational criterion.

    The surrogate criterion given at the end of the previous round—TVD = δ equivalent to “every sub-tier difference not positive”—has been verified across k=3/4/5/6 and multiple main-tier proportion differences, each tested on 400,000 random margins. Consistency is 100.00%, pass rate identical digit-by-digit. It requires no estimation of E, no knowledge of who is A and who is B, no dimension-by-dimension multiplication.

    The only two numbers needed for reporting are: main-tier proportion difference δ, and the remaining mass s_B on the larger main-tier side. The latter determines the contraction factor; the former is already in the report.

    This collides with an unresolved conflict edge within the research core.

    Our network has a grouping premise: labeling subjects via five-organ syndrome differentiation. This premise now hangs in conflict status, with evidence available on both sides. The pessimistic side: diagnostic consistency among TCM practitioners is “low” in open populations, with the original authors using the term “worrying” (PMID 30383398). The optimistic side comes from the same series: much of that “low” stems from statistical misuse; recalculating with Gwet AC2 yields linearly weighted consistency 0.80, chance-corrected falling between 0.67 and 0.73, and trainable improvement possible (PMID 19538100, PMID 29851495 specifically arguing AC2 is the appropriate statistic for free-margin data). Both sides refer to the same dataset; the disagreement lies in which coefficient is used. And this paper addresses precisely such matters: the looseness or tightness of a criterion may stem entirely from its assumed dimensionality, not the data itself. Thus, whether “low consistency means unreliable labels” holds after changing tier counts and coefficients cannot yet be treated as a settled conclusion.

    Evidence and Limitations

    Verified (pure mathematics, non-empirical): Under k tiers, the probability of “exactly not crossing the line” contracts as (δ/s_B)^(k−2); closed-form and Monte Carlo match digit-by-digit across eight parameter sets, each with two million samples; TVD ≤ δ is 100% equivalent to “all sub-tier differences ≤ 0” across k=3 to 6 globally; the three-tier 1+2E exactly equals the length ratio in one-dimensional cases, matching the volume ratio due to degenerate coincidence.

    Partially verified: Misjudgment rates of the tier-folding method, naive dimension-by-dimension method, and dimension-by-dimension (1+2E) approach have been computed within the simulation parameter grid, covering main-tier proportion difference ratios from 0.3 to 1.5 and tier counts from 3 to 6; regions outside this grid with extremely low consistency rates were not swept.

    Pending verification: In real adjudication studies, where do the main-tier proportion and main-tier difference for scales with four or more tiers actually fall? Current values assume uniform random allocation; real readers exhibit systematic preferences, making the distribution non-uniform and potentially shifting the contraction factor. Furthermore, “sample size exploding with more tiers” has never been written as a design constraint in published multi-tier consistency studies; it remains unclear whether such studies are therefore underpowered.

    L1 Network Status: L1 corresponds to an edge not yet found in the network (the effect of ordered scale tier count on the safety zone volume for consistency criteria). The edge table concept contains 61 entries (E-001 to E-061); none registers “scale tier count” as a variable. The closest, E-060 “Five-organ syndrome differentiation as grouping labels | premise | label reliability,” registers only “label reliability sufficient?” (KB-292/KB-317/KB-333/KB-334) in conflict status, without registering “the criterion for consistency coefficient itself changes shape with tier count.” Hence, this paper’s evidence level caps at C.

    Related Reading

    Evidence Level: C
    References: PMID 8834562, PMID 33853549, PMID 10769936, PMID 30869956, PMID 36050908, PMID 29851495, PMID 32020851

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


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

  • 那个「不知道」是一条随样本量移动的线:同一阶在两篇论文里不是同一个量,三种补法差两倍

    Evidence level: C

    逐阶表从「一串数」改成「一串数加一串状态」之后,跨研究合并这件事就不成立了——不是因为缺了数,而是因为「不知道」的边界本身由样本量决定,两篇论文里的同一阶从来就不是同一个量。

    一、先把「不知道」的来历说清

    上一轮把「不可判定」写进逐阶表,直觉上像补了一个缺失值。但它的产生机制和缺失完全不同:观测本身是有的,只是被一条判据线挡在外面。

    判据是 Rayleigh 型的:要认定某阶存在共享,该阶合向量长度须超过 2.448 除以根号 n,平方之后就是该阶份额超出 5.993 除以 n。于是「不知道」的边界不是生理常数,而是样本量的倒数——n 为 50 时线在 0.120,n 为 71 时在 0.084,n 为 300 时降到 0.020。

    同一个受试者、同一份生理,换一个样本量,状态就换了。这一点决定了后面全部的麻烦。

    二、三种补法各自在估什么

    把它们写成公式,问题立刻现形。丢掉「不知道」的那些研究,估的是「已知可判定的那部分的条件均值」;按判据线代入,估的是「真值与判据线取大之后的均值」;按零代入,估的是「真值乘以检出比例」。

    三者没有一个是「该阶份额的均值」。所以这不是三种做法有好有坏,而是三个不同的估计目标被塞进了同一个名字——那个名字叫「合并 S6」。

    已有方法学给出了更硬的处理路径。左删失数据的比较研究(PMID 25261453)在 10% 至 90% 删失、5 至 100 人、几何标准差 2 至 5 的条件下比较了极大似然、β 替代与 Kaplan-Meier 三法,结论是 β 替代总体不劣于另两者,且极大似然只在样本量不少于 30、删失不超 80%、变异较小(几何标准差 2 至 3)时才可靠;经典替代法(判据线除以 2、除以根号 2)的偏倚明显更大(PMID 20169489)。另有研究直接指出,简单替代在长尾分布下会系统性失真,须改用多重随机值填补(PMID 39141417)。换句话说:在三种坏答案里挑一个,本来就该被替换成按删失机制做似然。

    三、极差有多大

    以六阶真值 0.075 为例。样本量中位取 50(这恰是一致性研究连续终点的实测中位,PMID 36123642,四分位 25 至 100),检出率只有 31%,三种补法给出 0.095、0.155、0.029——极差 441%。中位取 71,检出率 45%,三者 0.094、0.130、0.042,极差 207%。中位取 119(分类终点的中位),极差降到 77%。

    要压到 20% 以内,六阶得把样本量中位提到 200 左右;若该阶真值更高(二阶 0.16),150 就够。

    这件事可以被一行式子概括:若不可判定者占 w,判据线是该阶真值的 f 倍,合并值相对真值就是「1 减 w 加 w 乘 f」。w 为一半、f 为 2.5 时,合并值直接变成真值的 1.75 倍。

    四、阶间斜率被压平还是放大

    把五阶一起算,真值从二阶 0.16 降到六阶 0.075,斜率每阶负 0.0210。丢缺失给出负 0.0219,相对真值只差 4%,几乎原样。按判据线代入给出负 0.0147,衰减三成——它把高阶压得没那么低了。按零代入给出负 0.0239,放大一成四。

    这里必须说一句反直觉的:若真值根本不随阶下降(各阶都 0.09),三种补法的斜率都落在正负 0.0004 以内,谁也没有凭空造出「高阶衰减」。所以问题不是伪造方向性结论,而是整体水平位移,加上阶间差异被压缩或放大。压平那三成最危险——它恰好削弱了「低阶共享高于高阶」这个本该成立的结论。

    五、留下的研究系统性是大样本

    再看被丢掉的是谁。15 份研究、样本量中位 71 的情形下,检出的那份平均 125 人,未检出的平均 33 人,差 3.79 倍;六阶差 3.48 倍。方向恒定,不随阶次翻转。

    后果是逐阶的「有效研究数」并不相等:二阶 11.5 份,三阶 9.7,四阶 7.9,五阶 7.1,六阶 6.8。高阶只剩不到一半的证据。而合并标准误反而不一定变宽——中位 71 时剔除后与全量的比是 1.06,中位 150 时是 0.90,精度看着更好了,因为留下的都是大样本。

    六、一半情形连排序都不一致

    三种补法即便都能报出一个数,也不代表它们指向同一个结论。重复 2000 次,三种处理同时指认二阶为最高阶的比例只有 51%。剩下那一半,至少有一种补法把最高阶指给了别人。

    这意味着「合并之后哪一阶最强」这个定性结论,在现有报告习惯下有一半概率不稳。

    七、判据与测法

    判据按四档:不可判定占比低于 20% 且三者极差低于 20%,可合并、三种补法任选;占比 20% 至 50%,必须三档并报、禁报单一合并值;占比高于 50%,只能逐研究报告、不做合并;三者对「哪阶最高」不一致的,该排序结论直接判不成立。

    测法上,正确路径不是在三种替代里挑,而是把判据线当作已知的删失边界去做似然(β 替代或极大似然),并对偏离随机缺失的假设做敏感性分析(PMID 30991455;网络 meta 中的信息性缺失可用贝叶斯模式混合模型处理,PMID 31134664)。任何跨研究合并都必须同时报:逐阶实际可判定研究数、判据线与该阶真值之比、以及删失的处理方式。三项缺一,那个合并值就不该进正文。

    证据与局限

    已验证:左删失数据的经典替代法(判据线除以 2、除以根号 2)存在已知偏倚,β 替代与极大似然更优,且极大似然有明确的适用边界(PMID 25261453、PMID 20169489);简单替代在长尾分布下失真,多重随机值填补更稳(PMID 39141417);一致性研究连续终点样本量中位仅 50(PMID 36123642);缺失结局的合并必须做偏离随机缺失假设的敏感性分析(PMID 30991455)。

    部分验证:该阶份额的检出下限等于 5.993 除以样本量,这是上一轮从判据常数 2.448 反推所得,属推导而非实测。

    待验证:真实脉搏谐波文献里「不可判定」的实际占比(本篇按 30% 至 50% 区间模拟);三种补法极差达 200% 这一量级在真实已发表数据上的实测值;以及不同量表合并时标准化均数差能否替代删失处理(PMID 32965803 指出量纲不同时只能标准化,但标准化本身不解决删失)。

    相关阅读

    证据等级: C
    参考文献: PMID 25261453, PMID 20169489, PMID 39141417, PMID 36123642, PMID 30991455, PMID 31134664, PMID 32965803

    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md 全表检索「合并、缺失、不可判定、状态、序数、删失、加权、汇总」无对应边;节点表仅在物理侧登记「相位、节律——周期性耦合的关键」一行(有对象、无层级、无状态),边表止于 E-061(呼吸节律 0.1 Hz 谱峰与 Mayer 波,状态 open)。本文证据等级按硬约束封顶为 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。

    That \”don’t know\” is a line that shifts with sample size: the same order represents different magnitudes across two papers, and three imputation methods differ by a factor of two.

    Evidence level: C

    After modifying the stepwise table from “a string of numbers” to “a string of numbers plus a string of states,” cross-study merging is no longer valid—not because numbers are missing, but because the boundary of “unknown” itself is determined by sample size. Consequently, the same order in two different papers never represents the same magnitude.

    I. Clarify the Origin of “Unknown” First

    In the previous iteration, “undecidable” was written into the stepwise table, intuitively resembling filling in a missing value. However, its generation mechanism differs fundamentally from missingness: observations exist, yet they fall outside a decision boundary.

    The criterion is Rayleigh-type: to establish that an order shares a common pattern, the length of its composite vector must exceed 2.448 divided by the square root of $n$; squared, this means the order’s share must exceed 5.993 divided by $n$. Thus, the boundary of “unknown” is not a physiological constant but the reciprocal of sample size—when $n=50$, the line sits at 0.120; when $n=71$, it drops to 0.084; and when $n=300$, it falls to 0.020.

    The same subject, the same physiological signal, with a different sample size, yields a different state. This fact underpins all subsequent complications.

    II. What Each of the Three Imputation Methods Estimates

    When written as formulas, the problem becomes evident. Studies that discard “unknown” estimate the conditional mean of the “known and decidable portion”; substituting along the criterion line estimates the mean of “true values truncated at the criterion line”; substituting zero estimates “true values multiplied by detection probability.”

    None of these is the mean of “the order’s share.” Therefore, this is not a matter of three methods being better or worse; rather, three distinct estimation targets are forced into the same name—”merged S6.”

    Existing methodology offers a more rigorous path. Comparative studies of left-censored data (PMID 25261453) evaluated maximum likelihood, β-substitution, and Kaplan-Meier under 10% to 90% censoring, 5 to 100 subjects, and geometric standard deviations of 2 to 5. The conclusion: β-substitution is overall not inferior to the other two, and maximum likelihood is reliable only when sample size ≥30, censoring ≤80%, and variability is low (geometric SD 2–3). Classical substitution methods (criterion line divided by 2, divided by $\sqrt{2}$) exhibit significantly greater bias (PMID 20169489). Another study directly states that simple substitution systematically distorts results under long-tail distributions, necessitating multiple imputation (PMID 39141417). In short: among three bad answers, one should be replaced by likelihood based on the censoring mechanism.

    III. How Large Is the Range?

    Take the sixth-order true value of 0.075 as an example. With median sample size $n=50$ (this matches the observed median of consistency study continuous endpoints, PMID 36123642; interquartile range 25–100), detection rate is only 31%. The three imputation methods yield 0.095, 0.155, and 0.029—a range of 441%. With median $n=71$, detection rate 45%, the three values are 0.094, 0.130, and 0.042, giving a range of 207%. With median $n=119$ (the median for categorical endpoints), the range drops to 77%.

    To reduce the range to within 20%, the sixth order must have a median sample size around 200; if this order’s true value is higher (second order at 0.16), 150 suffices.

    This can be summarized in one formula: if the undecidable portion accounts for $w$, and the criterion line is $f$ times the order’s true value, then the merged value relative to the true value equals “1 minus $w$ plus $w$ times $f$.” When $w=0.5$ and $f=2.5$, the merged value becomes 1.75 times the true value.

    IV. Are Inter-Order Slopes Flattened or Amplified?

    Including the fifth order, true values drop from second order 0.16 to sixth order 0.075, with a negative slope of −0.0210 per order. Discarding missing values yields −0.0219, differing from the true value by only 4%, nearly unchanged. Substituting along the criterion line yields −0.0147, reduced by 30%—it flattens higher orders. Substituting zero yields −0.0239, amplified by 14%.

    Here, a counterintuitive point must be stated: if true values do not decline across orders (all at 0.09), the slopes of all three imputation methods fall within ±0.0004—none artificially create “higher-order decay.” Thus, the problem is not fabricating directional conclusions, but overall level shifts combined with inter-order differences being compressed or amplified. Flattening that 30% is most dangerous—it precisely weakens the conclusion that “lower-order sharing exceeds higher-order sharing,” which should hold.

    V. Remaining Studies Are Systematically Large-Sample Biased

    Now consider which studies are discarded. In the scenario of 15 studies with median sample size 71, detected studies average 125 subjects, undetected studies average 33 subjects—a ratio of 3.79; for the sixth order, the ratio is 3.48. Direction remains constant and does not flip with order.

    The consequence is that the number of “effective studies” per step is unequal: second order 11.5, third order 9.7, fourth order 7.9, fifth order 7.1, sixth order 6.8—higher orders retain less than half the evidence. Yet merged standard error does not necessarily widen: at median 71, the ratio of post-removal to full quantity is 1.06; at median 150, it is 0.90—precision appears better because only large samples remain.

    VI. Half the Cases Even Fail to Agree on Ranking

    Even if all three imputation methods produce a number, they do not necessarily point to the same conclusion. Repeated 2,000 times, the proportion where all three simultaneously identify second order as highest is only 51%. In the remaining half, at least one imputation assigns highest order to someone else.

    This means the qualitative conclusion “which order is strongest after merging” has a 50% chance of instability under current reporting conventions.

    VII. Criteria and Measurement Methods

    Criteria are divided into four tiers: if undecidable proportion <20% and range of three methods <20%, merging is allowed with any imputation; if undecidable proportion 20–50%, all three must be reported and a single merged value forbidden; if undecidable proportion >50%, only individual reports are permitted, no merging; if the three methods disagree on “which order is highest,” that ranking conclusion is directly invalid.

    Regarding measurement, the correct path is not to choose among three substitutions, but to treat the criterion line as a known censoring boundary and perform likelihood estimation (β-substitution or maximum likelihood), while conducting sensitivity analysis for deviations from the random missing assumption (PMID 30991455; informative missingness in network meta-analysis can be handled via Bayesian pattern-mixture models, PMID 31134664). Any cross-study merging must simultaneously report: the actual number of decidable studies per step, the ratio of criterion line to true value for that order, and the handling of censoring. If any of these three is missing, the merged value should not appear in the main text.

    Evidence and Limitations

    Verified: Classical substitution for left-censored data (criterion line divided by 2, divided by $\sqrt{2}$) exhibits known bias; β-substitution and maximum likelihood are superior, with maximum likelihood having clear applicability boundaries (PMID 25261453, PMID 20169489); simple substitution distorts under long-tail distributions, making multiple imputation more stable (PMID 39141417); median sample size of consistency study continuous endpoints is only 50 (PMID 36123642); merging of missing outcomes must include sensitivity analysis for deviation from random missingness (PMID 30991455).

    Partially Verified: The detection lower bound for an order’s share equals 5.993 divided by sample size, derived from reversing the criterion constant 2.448 in the previous iteration—this is theoretical derivation, not empirical measurement.

    To Be Verified: The actual proportion of “undecidable” cases in real pulse harmonic literature (this paper simulates 30–50%); the observed range of 200% across three imputation methods on real published data; and whether standardized mean differences can substitute for censoring handling when merging different scales (PMID 32965803 notes that differing units require standardization, but standardization itself does not resolve censoring).

    Related Reading

    Evidence Level: C
    References: PMID 25261453, PMID 20169489, PMID 39141417, PMID 36123642, PMID 30991455, PMID 31134664, PMID 32965803

    L1 Network Status: This topic has not yet entered the L1 correspondence network—searching CORRESPONDENCE_NETWORK.md and EDGE_TABLE_science-yixue.md for “merge, missing, undecidable, state, ordinal, censored, weighted, summary” yields no corresponding edges; the node table registers only “phase, rhythm—key periodic coupling” (with objects, no hierarchy, no state), and the edge table stops at E-061 (respiratory rhythm 0.1 Hz spectral peak and Mayer wave, state open). This paper’s evidence level is capped at C by hard constraints.


    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.