Author: joe

  • tau 这个数测得出来,那条线判不过去:门槛附近九成五落在灰区,而判「过严」天生比判「成立」难

    Evidence level: C

    前面几篇把「以资深者为准能错多少」算到了小数点后一位,可那个数从头到尾建立在一个从没被测过的参数上:跨场景异质性 tau。这篇不谈它该多大,只回答一个更前置的问题——给它测一个值,这件事本身可不可做,测出来又能判到什么程度。

    一、那条线不是竖线,是斜线

    先把上一稿里的临界值摆正。错误率等于两倍的 p 乘一减 p,其中 p 是标准正态分布里 mu 除以 tau 那个位置的概率。令它等于百分之十五,解出来不是「tau 小于等于三点五九个百分点」,而是mu 除以 tau 大于等于一点三九四。

    这个换算一写出来,原来那句话的问题就露了:三点五九这个数字,只在 mu 恰好等于五个百分点时成立。mu 换成两个百分点,门槛掉到一点四三;换成十个百分点,门槛抬到七点一七。同一句「tau 小于三点五九故禁令过严」,在 mu 等于八个百分点时对应的真实错误率只有百分之二点六,在 mu 等于两个百分点时却是百分之四十一。

    所以回溯不能只交一个 tau。每一组数据必须带上它自己的 mu,否则那个绝对门槛是借来的量纲——和上一篇在 logit 换算上踩的是同一个坑。

    二、三十组够测 tau 吗:够,中位数几乎不偏

    假设每组三十例、单次判读噪声八个百分点,用经典矩估计跑三千次模拟。真 tau 为二时,点估计的中位数是一点九四;为三时二点九四;为五时四点九二。偏差小到可以忽略。

    区间估计也在名义水平上:三种真值下覆盖率都在百分之九十四点八到九十五点一之间。方法学上有明确背书——九种异质性方差估计量的系统比较推荐限制最大似然,并明确指出矩估计在小样本、稀有结局时系统性偏低(PMID 30067315,Langan 等,Res Synth Methods 2019);十六种方法的比较则直接建议把区间和点估计一起报(PMID 26332144,Veroniki 等)。Q 剖面法做区间本身是标准做法,但 Q 统计量一族的方法在实践条件下覆盖率会明显低于名义值(PMID 30589219,van Aert 等),所以这里的百分之九十五是乐观上限。

    结论:测 tau 不难。三十组、每组三十例,够用了。

    三、难的是「判过没过线」,而良性和有害不对称

    这才是本篇真正的发现。

    对着三点五九这条线做判定,如果要求区间整体落在门槛一侧才算判出,结果是这样的:真 tau 等于二(禁令其实过严,错误率仅百分之一点二),三十组时判对只有百分之七十三点八,四分之一被判成「禁令成立」。真 tau 等于五(禁令确实成立,错误率百分之二十六点七),同样三十组,判对百分之九十六点一,误判只剩百分之三点九。

    一边七十四,一边九十六。原因很朴素:判「过严」要用区间的上界去低于门槛,而上界天生就宽——三十组时上界的中位数是点估计的一点四四倍;判「成立」只需要点估计越过门槛。

    也就是说,这套流程对「维持禁令」宽松,对「推翻禁令」严苛。举证责任挂在想推翻的那一方身上,而且是无意识挂上去的。 这跟统计里常见的保守方向相反:通常保守是「不敢说有效」,这里是「不敢说没问题」。

    更要紧的是,门槛正上方那段几乎测不出来。真 tau 恰好等于三点五九时,百分之九十四点九的样本落在灰区——既不能说它过严,也不能说它成立。而这个区间恰恰是唯一真正需要答案的地方。

    四、加病例几乎没用,瓶颈是组数

    直觉会认为多收病例就能把 tau 测准。算一下就知道不行。

    固定三十组,每组例数从三十加到一千九百二十(六十四倍),mu 的估计标准误只从零点七六降到零点六六个百分点。原因是 tau 平方在总方差里的份额从百分之七十五涨到百分之九十九点五——噪声项被压干了,剩下的全是真异质性,而它只能靠增加组数来平均掉。

    要把「错误率区间宽度」压到十个百份点以内,每组三十例时需要八十组;每组六十例时需要五十组。要判出「良性」达到九成五把握,真 tau 等于二需要八十组,等于三需要五百组,等于三点四需要八百组以上。

    越接近门槛,越判不出。 这不是精度问题,是这套判据的结构问题。

    五、小样本还会把良性伪装成完美

    还有一层。经典矩估计在小 K 下会被零截断——真 tau 等于一、K 等于十时,四成样本的 tau 被压成零。零意味着「各组完全一致、没有任何跨场景波动」,读起来是最理想的状态。

    于是出现一种反向失真:真正最良性、波动最小的那批数据,有相当比例被报成「波动为零」。而零这个值在业务上会直接被读成「这条禁令纯属多余」。

    六、能给的与不能给的

    能给的:一个三十到八十组、每组三十例的回溯,可以把 tau 从假想参数变成带区间的实测值,中位数基本无偏。这是可做的。

    不能给的:用同一个回溯去裁决「禁令该留该废」。门槛附近判不出,良性一侧四分之一错判,而加病例解决不了。

    正确用法:不问「tau 有没有超过三点五九」,改问「mu 除以 tau 的区间在哪里」,并且报告这个比值而不是 tau 的绝对值。比值在八十组时区间仍是一点零六到一点八六,对应错误率百分之二十五到百分之六——即便如此宽,它至少不借量纲、不随 mu 漂移。

    证据与局限

    已验证: 矩估计在中小 K 下点估计近似无偏、Q 剖面区间覆盖率接近名义值,这两点在本篇模拟与方法学文献(PMID 30067315、PMID 26332144)中一致;限制最大似然优于矩估计是共识结论。

    部分验证: 「判良性比判有害难」这一不对称在本篇模拟中稳定复现(三十组时七十四对九十六),但它依赖「区间整体落在一侧才算判出」这条判定规则。换用其他规则(例如只看点估计)不对称会消失,代价是失去控制。判定规则本身没有外部标准可依,属本篇设定。

    待验证: ① mu 与 tau 在真实判读数据里是否独立——若 mu 与 tau 相关(严厉的中心波动也大),比值判据的分布会变形,本篇按独立生成,未验;② 三十组的历史数据在文献里是否找得到——多数一致性研究报告的是合并 kappa 而非分组原始差,回溯可能根本凑不齐 K;③ 本模拟假设每组等例数、等噪声,真实数据里例数与噪声同时变化,加权方式需另行检验。

    相关阅读

    证据等级: C
    参考文献: PMID 30067315, PMID 26332144, PMID 30589219, PMID 28122727
    L1 网络状态:CORRESPONDENCE_NETWORK.md(1956 行)逐词统计:「异质」0 处、「随机效应」0 处、「tau」0 处、「错误率」0 处、「kappa」0 处、「方差」1 处(第 1694 行 E-005 声音韵律条目,指基频方差这一声学量,与跨场景方差成分无关)、「偏倚」1 处(第 265 行 KB-338 自陈叙述性综述无偏倚评估)、「评定者」4 处、「评分者」1 处、「判读」1 处(第 1832 行 E-061 驱动方向判读须跨两族方法),全部落在 RSA 读数规格与穴位对照臂清单层。EDGE_TABLE_science-yixue.md(63 行)grep「kappa/信度/一致/方差/误差/偏倚/校正/异质/随机效应」全部 0 命中。该主题最接近的边仍是 E-060 标签前提(五脏辨证作分组标签|前提|标签信度,状态 conflict 未裁决)——KB-317 记两处错配、KB-333 记序数分类下系数选择与估计精度须分层裁决、KB-334 要求 AC2 与患病率分报;该边只登记「标签信度够不够用」,未登记「跨场景异质性有多大、以及它能否被实测出来」。该主题尚未进入 L1 对应网络,已登记至 science-yixue-l1-gaps.md(「测量一致性」簇第 60 条)。


    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。

    Tau is measurable, but that line cannot be crossed: ninety-five percent of the threshold falls in the gray zone, and judging \”too strict\” is inherently harder than judging \”esta…

    Evidence level: C

    Previous articles calculated “how much error is acceptable for experts” to one decimal place. However, that number relies entirely on an untested parameter: cross-scenario heterogeneity tau. This paper does not determine its magnitude; it answers a more fundamental question: Can we measure tau at all, and if so, what can the result tell us?

    1. The line is not vertical; it is slanted

    First, correct the threshold from the previous draft. The error rate equals twice p times (1 − p), where p is the probability at the position of mu divided by tau in the standard normal distribution. Setting this equal to 15% yields not “tau ≤ 3.59%” but mu / tau ≥ 1.394.

    This conversion exposes the flaw in the previous statement: the figure 3.59 holds only when mu equals exactly 5%. If mu is reduced to 2%, the threshold drops to 1.43; if mu increases to 10%, the threshold rises to 7.17. The claim “tau ≤ 3.59, so the ban is overly strict” is therefore misleading. When mu = 8%, the true error rate corresponding to this statement is only 2.6%; when mu = 2%, it jumps to 41%.

    Thus, one cannot simply report tau. Each dataset must include its own mu, otherwise the absolute threshold borrows units from another context—falling into the same pit as the logit conversion in the previous paper.

    2. Are thirty groups enough to estimate tau? Yes, the median is nearly unbiased

    Assuming 30 cases per group and single-reading noise of 8%, running 3,000 simulations via classical moment estimation yields:
    – True tau = 2 → median point estimate ≈ 1.94
    – True tau = 3 → median point estimate ≈ 2.94
    – True tau = 5 → median point estimate ≈ 4.92

    The bias is negligible.

    Interval estimation also performs at nominal levels: coverage across all three true values ranges from 94.8% to 95.1%. Methodologically, this is well-supported: systematic comparisons of nine heterogeneity variance estimators recommend restricted maximum likelihood over moment estimation, noting that the latter tends to be systematically biased low in small samples and rare outcomes (PMID 30067315, Langan et al., Res Synth Methods 2019); comparisons of sixteen methods further advise reporting both intervals and point estimates together (PMID 26332144, Veroniki et al.). While the Q-profile method for intervals is standard, Q-statistic-based methods yield coverage noticeably lower than nominal levels under practical conditions (PMID 30589219, van Aert et al.), making the 95% here an optimistic upper bound.

    Conclusion: Estimating tau is straightforward. Thirty groups of thirty cases each are sufficient.

    3. The real difficulty lies in deciding “over or under the line,” and benign vs. harmful asymmetry

    This is the paper’s core finding.

    When judging against the 3.59 threshold, requiring the entire interval to fall on one side to declare a decision yields:
    – True tau = 2 (ban actually too strict, error rate only 1.2%): correctly identified in 73.8% of cases, but one-quarter are misclassified as “ban holds.”
    – True tau = 5 (ban truly valid, error rate 26.7%): correctly identified in 96.1% of cases, with only 3.9% misclassification.

    The asymmetry is stark: 74% vs. 96%. The reason is straightforward: declaring “too strict” requires the upper bound to be below the threshold, and upper bounds are inherently wider—median upper bounds are 1.44 times the point estimate with 30 groups. Declaring “valid” merely requires the point estimate to exceed the threshold.

    Thus, this procedure is lenient toward “maintaining the ban” but strict toward “overturning it.” The burden of proof falls on the side seeking to overturn, and it is imposed unconsciously. This contrasts with common statistical conservatism, which usually means “not claiming effectiveness”; here, it means “not claiming no problem.”

    Crucially, the region just above the threshold is nearly impossible to resolve. When true tau equals exactly 3.59, 94.9% of samples fall into a gray zone—neither clearly too strict nor clearly valid. This interval is precisely where answers matter most.

    4. Adding cases helps little; the bottleneck is the number of groups

    Intuition suggests more cases improve tau estimation. Calculation shows otherwise.

    Fixing 30 groups and increasing cases per group from 30 to 1,920 (64-fold), the standard error of mu drops only from 0.76 to 0.66 percentage points. This occurs because tau² accounts for 75% of total variance initially, rising to 99.5%—noise is suppressed, leaving only true heterogeneity, which can only be averaged out by increasing the number of groups.

    To shrink the “error rate interval width” below 10 percentage points:
    – 30 cases per group → need 80 groups
    – 60 cases per group → need 50 groups

    To achieve 95% confidence in declaring “benign”:
    – True tau = 2 → requires 80 groups
    – True tau = 3 → requires 500 groups
    – True tau = 3.4 → requires more than 800 groups

    The closer to the threshold, the harder it is to decide. This is not a precision issue but a structural limitation of the decision rule.

    5. Small samples can disguise benign cases as perfect

    A further complication arises from small K. Classical moment estimation is truncated at zero when K is small: with true tau = 1 and K = 10, 40% of samples have tau estimated as zero. Zero implies “complete agreement across groups, no cross-scenario fluctuation”—the ideal state.

    This creates reverse distortion: the most benign, least variable data are disproportionately reported as “zero fluctuation.” Practically, zero is read as “this ban is entirely unnecessary.”

    6. What can be given and what cannot

    What can be given: A retrospective of 30 to 80 groups, each with 30 cases, can convert tau from a hypothetical parameter into an interval-based measured value, with median estimates nearly unbiased. This is feasible.

    What cannot be given: Using the same retrospective to decide “should the ban remain or be abolished.” Near the threshold, decisions are unattainable; benign cases are misclassified in one-quarter of instances; adding cases does not resolve this.

    Correct usage: Do not ask “Did tau exceed 3.59?” Instead, ask “Where does the interval of mu / tau lie?” and report this ratio rather than the absolute value of tau. With 80 groups, the ratio interval ranges from 1.06 to 1.86, corresponding to error rates of 25% to 6%. Even this wide interval avoids borrowing units and drifting with mu.

    Evidence and Limitations

    Verified: Moment estimation yields nearly unbiased point estimates for small-to-moderate K, and Q-profile intervals achieve coverage close to nominal levels; these findings are consistent with this paper’s simulations and methodological literature (PMID 30067315, PMID 26332144). Restricted maximum likelihood outperforms moment estimation, which is consensus.

    Partially verified: The asymmetry that “benign cases are harder to identify than harmful ones” is stably reproduced in this simulation (74 vs. 96 with 30 groups), but it depends on the decision rule requiring the entire interval to fall on one side. Alternative rules (e.g., using only point estimates) eliminate asymmetry at the cost of control. Since no external standard exists for the decision rule itself, this is a design choice of this paper.

    To be verified:
    1. Whether mu and tau are independent in real judgment data—if they are correlated (e.g., severe central fluctuation implies large mu), the distribution of the ratio-based criterion would distort; this paper assumes independence.
    2. Whether historical data with 30 groups can be found in literature—most consistency reports aggregate kappa rather than reporting group-level differences, making reconstruction of K unlikely.
    3. Whether the simulation assumption of equal cases and noise per group holds—real data often vary both, requiring separate weighting validation.

    Related Reading

    Evidence level: C
    References: PMID 30067315, PMID 26332144, PMID 30589219, PMID 28122727
    L1 Network Status: CORRESPONDENCE_NETWORK.md (1,956 lines) word-count: “heterogeneous” 0 times, “random effects” 0 times, “tau” 0 times, “error rate” 0 times, “kappa” 0 times, “variance” 1 time (line 1,694, E-005 sound rhythm entry, referring to fundamental frequency variance, an acoustic measure unrelated to cross-scenario variance components), “bias” 1 time (line 265, KB-338 self-reported narrative review without bias assessment), “rater” 4 times, “scorer” 1 time, “judgment” 1 time (line 1,832, E-061 driving direction judgment requires cross-family methods), all falling within RSA reading specifications and acupoint control arm lists. EDGE_TABLE_science-yixue.md (63 lines) grep for “kappa / reliability / consistency / variance / error / bias / calibration / heterogeneous / random effects” yields zero hits. The closest edge is E-060 label premise (organ differentiation as grouping labels | premise | label reliability, status conflict unresolved)—KB-317 records two mismatches, KB-333 notes ordinal classification requires coefficient selection and estimation precision to be tiered, KB-334 demands AC2 and prevalence reporting separately; this edge only registers “is label reliability sufficient,” not “how large is cross-scenario heterogeneity and can it be measured.” This topic has not yet entered the L1 corresponding network, registered instead in science-yixue-l1-gaps.md (cluster 60 of “measurement consistency”).


    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.

  • 偏科比偏得多更伤人:残差是常数还是一截阶梯,取决于偏倚散不散、而不取决于偏得多大

    Evidence level: C

    上一篇说「校正之后剩下的偏差,精确等于资深者自己的偏倚」。那句话还有个前提被藏起来了:这口偏倚得是一个常数。一旦它随病例亚型上下浮动,你减掉的就不是一个数,而是一个平均——剩下的东西从一口平底锅变成一截台阶。

    一、先把「残差是常数」这句话拆开

    两位判读者看同一批病例:资深者 A 读出「真值 + a」,新人 B 读出「真值 + b」。把 B 整体减去两人的平均差,剩下的是 a。这是代数,没有争议。

    问题在于 a 被当成了跨病例不变的常数。这个假定在判读研究里几乎站不住——42306452 号研究用多面 Rasch 的差异性评分者功能分析 69 位评分者对 3875 份申请的判读,发现「是否与申请者同性别、同族群背景」会改变同一位评分者的严厉度,偏移量在 0.46 到 1.74 logit 之间;35754620 号研究进一步把这个框架扩成「严厉度与居中倾向同时偏科」的双通道模型,并证明双通道偏科会实质损害测量质量。换言之,一位评分者不是「整体偏严三分」,而是「对这一类偏严、对那一类偏宽」。

    若 a 随亚型 k 浮动、b 不随,则校正后的残差是 a_k,不是一个数。它的均方值等于「平均偏倚的平方加上亚型间方差」。前半截随样本量下降,后半截不降——你多测多少例都压不掉它。

    二、开关不是偏得多大,而是偏得多散

    把「亚型间方差占偏倚总方差的比例」当判据,会撞上一个尴尬:平均偏倚趋于零时这个比例必然趋于百分之百,于是「偏科最严重」的判断恰好落在「校正等于白校正」的最良性情形上。判据用错了。

    换成无量纲的比值——亚型间标准差除以平均偏倚的绝对值——界线立刻干净:这个比值小于三分之一,离散只占总方差一成以下,单常数近似可用;到三分之二,离散占到三成,「校正后残差为常数」这句话必须改成「按亚型分段」;比值超过一,均方残差已经一半以上来自离散,减一个统一数已经没有意义。

    这条换算和偏倚的绝对大小无关,只和「散不散」有关。

    三、同一份数据里,这个比例已经悬殊到十几倍

    42306452 号研究把两类偏倚都报了出来:跨组平均的严厉度偏移只有负 0.12 与负 0.07 logit,而分群偏移是 0.46 与 1.74 logit。前者是后者的百分之七到十五。

    照上面的换算,比值落在七到十五之间,离散占比超过百分之九十八。这不是边缘超标,是单常数近似彻底失效——连一个能统一减掉的数都拿不出来。该研究还给了另一个方向的证据:超过一半的个体评分者在性别维度上呈现可观的偏科,三成左右在族群维度上如此。也就是说这不是少数人的毛病,而是多数人都有、只是方向各自不同——平均起来互相抵消,看起来「整体没偏」。

    「整体偏小」最危险的地方就在这里:它由一堆方向相反的大偏移抵消而成,剩下的那个小数字最不具代表性。

    四、阶梯看得见吗:假台阶比真台阶更容易出现

    假设每个亚型只有三十例、单次判读噪声八个百分点。即使真实偏科为零,仅仅因为抽样,实测到的亚型间方差期望值就有四点三平方百分点。此时「假台阶」占到百分之百。

    真实离散要多大才压得住这个地板?标准差一,假台阶八成;二,五成;三,三成;四点六,一成七;八,才降到百分之六。要让假台阶压到一成以下,每个亚型得有一百一十到一千一百例,取决于真实离散是一到八。

    齐性检验的处境同样难看。三个亚型各三十例、真偏科为负二、零、正二时,检出率只有两成二;要到各一百二十例才六成九。真偏科翻到三,三十例时四成四、一百二十例时九成七。37810544 号研究在模拟里也撞到同一堵墙:稀疏设计下差异性评分者功能的对比估计幅度系统性低于设定真值,甚至「明明注入了偏科,与评分者交互的方差成分也不一定能读出来」。25546697 号研究做量表跨评定者可比性检验时,也是先控住评分者效应才敢谈可比。

    所以现在能写的话只有一句:多数判读研究的样本量,对「有没有偏科」这个问题给不出信息;写「未检出差异」是把无信息当成了阴性证据。

    五、还有个更隐蔽的:代入平均值会错到五倍

    若关心的是「隔多久判错一次」,错误率等于二倍 p 乘一减 p,p 是平均偏倚除以跨场景异质性再取正态分布函数。这个式子是凹的,代入平均值必然系统偏差,而且方向随平均偏倚离零的距离翻转。

    算一遍:跨场景异质性三个百分点时,偏倚在二、六、十一、十六之间浮动(均值六点四),逐亚型加权算出的错误率是一成六,代入平均值算出的是三点二——差了五倍,代入法把风险说小了。反过来,偏倚在负四、负一、正三、正四之间(均值负零点九),加权算出三成九,代入算出四成九——这次代入法把风险说大了。

    异质性涨到五个百分点,倍数差收到一点四倍;涨到八以上,两种算法几乎同值(一比一)。原因是凹性的曲率被摊平了:异质性大时所有亚型都被推到「接近五成」的天花板,互相抹平。于是有个反讽的结论——跨场景越一致(异质性越小),这套算法越「好用」,也越不准。

    证据与局限

    已验证:以资深者为准的校正,在偏倚恒定前提下残差精确等于资深者自身偏倚(代数恒等式)。偏科现象在真实判读数据里确实存在且普遍:42306452 号研究中过半评分者在性别维度呈现可观偏科,双通道偏科会损害测量质量(35754620)。稀疏设计会系统性压低偏科的检出灵敏度(37810544)。

    部分验证:本文的三档判据(三分之一、三分之二、一倍)来自方差分解的恒等式换算,数学上是精确的;但它们对应的「可用/须分段/失效」边界,目前只是量级判断,尚无实测数据校准。「假台阶」与检验力的具体数字基于单次判读噪声八个百分点、亚型各三十例这组设定,噪声或例数一变,整张表都要重算。

    待验证:亚型间偏倚标准差在实际判读任务里到底多大,没有任何一篇文献报过。42306452 给出的是 logit 尺度的分群偏移,不是百分点;两类量纲能否直接换算、换算后比值落在哪一档,都是空白。此外,「偏科是否随时间稳定」也未测——若它本身是随机漂移,那么分段标定同样会被噪声吃掉。

    相关阅读

    证据等级: C
    参考文献: PMID 42306452, PMID 35754620, PMID 37810544, PMID 28328280, PMID 25546697, PMID 33853549, PMID 30383398

    L1 网络状态:该主题尚未进入 L1 对应网络。EDGE_TABLE_science-yixue.md 全 63 行中「kappa/信度/一致/方差/误差/偏倚/校正/亚型/分层」全部 0 命中;CORRESPONDENCE_NETWORK.md 中最接近的仍是第 223 行「E-060 标签前提」(五脏辨证作分组标签|前提|标签信度|状态「冲突(未裁决)」)——该边只登记「标签信度够不够用」,未登记「偏倚随亚型变化时残差从常数变阶梯、以及该阶梯能否被现有样本量看见」。已登记至 L1 缺口表。

    证据等级说明:A 表示有 PubMed 直接支撑该结论,B 表示科学类比或间接证据,C 表示传统文化共识或初步探索。本文标 C,是因为该边未进入 L1 网络,且核心判据(三档比值)为代数换算而非实测校准。


    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。

    Bias hurts more when it is biased: whether residuals are constant or a step depends on how scattered the bias is, not on how much it is biased.

    Evidence level: C

    The previous post stated that “the residual error after correction precisely equals the senior rater’s own bias.” That statement had a hidden premise: this bias is a constant. Once it fluctuates with case subtypes, what you subtract is no longer a single number but an average—the remainder transforms from one flat pan into a stepped ladder.

    I. First, Deconstruct “Residual Error Is Constant”

    Two raters assess the same cases: Senior A reads “True Value + a”, Novice B reads “True Value + b”. Subtracting B’s overall difference from the average leaves a. This is algebra, undisputed.

    The problem lies in treating a as a constant across cases. This assumption is nearly untenable in rater studies—Study 42306452 used multi-faceted Rasch differential rater function analysis on 69 raters scoring 3875 applications, finding that “whether the applicant shares gender or ethnic background” alters a single rater’s severity, with offsets ranging from 0.46 to 1.74 logits; Study 35754620 further expanded this framework into a dual-channel model of “severity and centrality bias simultaneously,” proving that dual-channel bias substantially degrades measurement quality. In other words, a rater is not “overall three points stricter” but “stricter for this group, lenient for that.”

    If a fluctuates with subtype k while b remains fixed, the corrected residual becomes a_k, not a single number. Its mean square equals “the square of average bias plus variance among subtypes.” The first half decreases with sample size, the second does not—no matter how many cases you add, it cannot be suppressed.

    II. The Switch Is Not How Much Bias, But How Scattered

    Using “the proportion of variance due to inter-subtype differences” as the criterion hits an awkwardness: when average bias approaches zero, this proportion necessarily approaches 100%, so the judgment of “most severe bias” coincides with the most benign case where “correction equals white correction.” The criterion is misapplied.

    Switching to a dimensionless ratio—inter-subtype standard deviation divided by absolute average bias—the boundary becomes clean: this ratio below one-third means dispersion accounts for less than 10% of total variance, single-constant approximation holds; at two-thirds, dispersion reaches 30%, “residual after correction is constant” must change to “segmented by subtype”; exceeding one, more than half of mean square residual stems from dispersion, subtracting a uniform number is meaningless.

    This conversion depends solely on “how scattered,” not on the absolute magnitude of bias.

    III. Within the Same Data, This Ratio Varies by Over Tenfold

    Study 42306452 reported both types of bias: cross-group average severity offset was only -0.12 and -0.07 logits, while subgroup offsets were 0.46 and 1.74 logits—the former is 7% to 15% of the latter.

    According to the above conversion, the ratio falls between seven and fifteen, with dispersion accounting for over 98%. This is not marginal exceedance but complete failure of single-constant approximation—no unified number can be subtracted. The study also provided evidence in the opposite direction: over half of individual raters showed significant gender-based bias, about one-third ethnic-based. Thus this is not a minority issue but widespread among most, just with differing directions—averaging cancels them out, appearing “overall unbiased.”

    The danger of “overall slightly biased” lies here: it results from cancellation of large opposing shifts, leaving the small residual least representative.

    IV. Can We See the Steps? False Steps Are More Likely Than True Ones

    Assume each subtype has only 30 cases and single-rater noise is 8 percentage points. Even if true bias is zero, merely due to sampling, the expected inter-subtype variance is 4.3 squared percentage points. At this point, “false steps” account for 100%.

    How large must true dispersion be to suppress this floor? Standard deviation of 1 yields 80% false steps; 2, 50%; 3, 30%; 4.6, 17%; 8, down to 6%. To reduce false steps below 10%, each subtype needs 110 to 1,100 cases, depending on true dispersion ranging from 1 to 8.

    Homogeneity testing faces similar difficulties. With three subtypes each of 30 cases and true biases of -2, 0, +2, detection rate is only 22%; reaching 69% requires 120 cases each. When true bias doubles to ±3, detection reaches 44% at 30 cases and 97% at 120. Study 37810544 encountered the same wall in simulation: under sparse design, differential rater function comparison estimates systematically fall below true values, even though “bias was injected, the variance component interacting with raters may still not be detectable.” Study 25546697, when testing scale cross-rater comparability, first controlled rater effects before claiming comparability.

    Thus, the only statement we can write is: most rater studies lack sufficient sample size to inform “whether bias exists”; declaring “no difference detected” mistakes null information for negative evidence.

    V. Another Hidden Issue: Substituting Averages Can Err by Fivefold

    If we care about “how often errors occur,” error rate equals 2p(1−p), where p is the average bias divided by cross-scenario heterogeneity, transformed via normal distribution function. This formula is concave; substituting averages inevitably introduces systematic error, with direction reversing based on distance from zero.

    Calculating: when cross-scenario heterogeneity is 3 percentage points, bias fluctuates between 2, 6, 11, and 16 (mean 6.4). Weighted calculation by subtype yields error rate of 1.6%, while substituting average gives 3.2%—five times higher, averaging underestimates risk. Conversely, when bias ranges from -4, -1, +3, +4 (mean -0.9), weighted calculation gives 3.9%, substituted gives 4.9%—this time averaging overestimates risk.

    When heterogeneity rises to 5 percentage points, the ratio drops to 1.4; exceeding 8, both methods converge (1:1). The reason lies in concavity flattening: high heterogeneity pushes all subtypes toward the “near 50%” ceiling, smoothing each other out. Thus comes an ironic conclusion—the more consistent across scenarios (lower heterogeneity), the more “useful” yet inaccurate this algorithm becomes.

    Evidence and Limitations

    Verified: Correction based on senior raters, under constant bias assumption, yields residual error precisely equal to the senior rater’s own bias (algebraic identity). Subtype bias indeed exists and is widespread in real rater data: Study 42306452 found over half of raters exhibited significant gender-based bias; dual-channel bias degrades measurement quality (35754620). Sparse design systematically reduces detection sensitivity of subtype bias (37810544).

    Partially Verified: The three-tier criteria (one-third, two-thirds, one) derive from variance decomposition algebraic conversion, mathematically exact; however, the boundaries for “usable/segmented/failing” remain magnitude estimates without empirical calibration. Specific numbers for “false steps” and test power rely on the setting of 8 percentage point single-rater noise and 30 cases per subtype; any change in noise or case count requires recalculation of the entire table.

    To Be Verified: No literature reports the actual standard deviation of inter-subtype bias in real rater tasks. Study 42306452 provides logit-scale subgroup offsets, not percentage points; whether these units can be directly converted and which tier the resulting ratio falls into remains unknown. Furthermore, “whether bias stabilizes over time” has not been tested—if it itself drifts randomly, segmentation calibration would also be consumed by noise.

    Further Reading

    Evidence Level: C
    References: PMID 42306452, PMID 35754620, PMID 37810544, PMID 28328280, PMID 25546697, PMID 33853549, PMID 30383398

    L1 Network Status: This topic has not yet entered the L1 corresponding network. In EDGE_TABLE_science-yixue.md (63 lines total), all terms “kappa/reliability/consistency/variance/error/bias/correction/subtype/layering” have zero hits; the closest in CORRESPONDENCE_NETWORK.md is line 223 “E-060 label premise” (organ differentiation as grouping labels | premise | label reliability | status “conflict (unresolved)”)—this edge only records “whether label reliability suffices,” not “when bias varies by subtype, residual shifts from constant to stepped, and whether existing sample size can detect this step.” Registered in L1 gap table.

    Evidence Level Explanation: A indicates direct PubMed support for the conclusion; B denotes scientific analogy or indirect evidence; C represents traditional consensus or preliminary exploration. This article is marked C because the edge has not entered the L1 network, and the core criterion (three-tier ratio) stems from algebraic conversion rather than empirical calibration.


    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.

  • 共享校准病例买到的不是精度,是「同一个真值」:曲线九成的好处压在最后那几个百分点上

    Evidence level: C

    上一篇给了一条便宜处方:多中心别各自招校准病例了,共享一部分。听着合理,但从没人算过中间态。算完它塌了:共享三成校准病例,在十几倍的膨胀面前几乎什么都没买到。而且你以为共享的是病例,其实共享的只是那个人的真值——误差还各自带着。

    正文

    上一篇(#1422)的结论是:校准十人、各抽各的,两人之差的精度要付十五倍代价,主项是「校准抽多少人」。它结尾留了条省钱出路——多中心共享一部分校准病例。这条在代数上没错,但中间态的形状没人算过。

    一、膨胀的平方里,重叠只买到一个线性折扣

    两位读者各取 n_cal 例校准、交集 m 例,重叠率 r 等于 m 除以 n_cal。两种误差来源要分开。

    若锚误差在重叠病例上是同一份(同设备、同一位资深者看同一批片子),它在两位读者位置估计的减法里被完整抵消,闭式解是:膨胀等于根号下 1 加 K 乘 1 减 r,K 等于重复次数乘样本数乘 1 加锚噪平方比,再除以校准人数乘读者噪平方。

    若锚误差每次都独立(同批病例、不同设备重测),重叠只让真值那部分对消,锚项原封不动:膨胀等于根号下 1 加 B 乘 1 减 r 再加 A,其中 A 是锚项、B 是真值项。

    两式与蒙特卡洛逐格对照,偏差在 1% 内。关键在第一项:重叠只出现在 1 减 r 这个乘子上,而它乘的是几十到几百的基数。 折扣是真的,基数也是巨大的。

    二、三成重叠,在十五倍的坑里只填掉两倍半

    取样本 240、读者噪 0.5、每人重复两次、锚噪 0.5。校准十人时 K 等于 240,无重叠膨胀 15.5 倍。

    重叠从 0 走到 0.3,膨胀从 15.5 降到 13.0。省掉 2.5 倍是真的,剩下的 13 倍也是真的。要压到两倍以内,需 r 等于 0.9875——校准十人里九点九个都得是同一批人,那已经等于不共享。

    校准加到三十、六十、一百二十人,门槛分别是 0.9625、0.925、0.850。只有校准人数接近全样本(240 人,K 等于 10)时,门槛才掉到 0.700,共享三成才有意义。

    「共享」不是「加人」的替代品,是它的下游。 校准人数没到位时谈共享,是在一个还没开始的故事里省钱。

    三、曲线的形状是反的:省的地方在最后,不在开头

    按线性直觉,重叠该「开头省得多、后面省得少」。实际相反——膨胀对 r 的导数分母里带着根号,r 越接近 1 分母越小、斜率越陡。

    校准十人这一格:重叠从 0 到 0.1,膨胀降 0.8 倍;从 0.9 到 1.0,降 4.0 倍。最后那一成买到的,是开头那一成的五倍。

    含义很硬:「共享三成」这种折中方案,恰好落在整条曲线最平的一段。 想要真收益就得几乎全共享;做不到,不如把钱花在抬高校准人数上。

    四、两种宇宙在零重叠时完全一样

    两个式子在最该被注意的地方露了一手:r 等于 0 时给同一个数 15.5 倍;r 等于 1 时,一个给 1.00 倍、一个给 7.00 倍。

    意思是:做重叠之前,你无法从数据判断自己活在哪个宇宙里。 而这条干预的全部价值,就是这两个宇宙在满重叠处的差。同源误差下全共享可做到零代价;独立误差下全共享仍留 7 倍地板。

    这个地板还不随样本量下降:样本从 60 加到 240 再到 960,实测 4.28、4.09、4.11。加人、加读者、加病例全部无效——它不是抽样误差。

    所以「共享校准病例」必须先答一个前问题:共享的到底是什么。 共享一个病例,只共享了那个人的真值;要让误差也对消,必须共享同一次测量事件——同设备、同判读者、同一时段。多中心各自用自己的设备测同一批人,那叫共享真值,不叫共享误差。

    这也解释了实验室里那条老经验:脂质组学跨平台跨实验室的系统偏差靠「共同参考样本」压下去,而它之所以有效,是因为该样本在每个平台上被同一套流程反复测——共享的是测量事件,不是样本本身。

    五、什么时候重叠才是一条路

    对独立误差源,全重叠残余为根号下 1 加 A。压到两倍内需 A 不超过 3,即校准人数至少为锚噪平方乘重复次数乘样本数除以三倍读者噪平方。本参数下是 160 人,占全样本 67%。

    三分之二,是重叠从「聊胜于无」变成「可替代加人」的分水岭。 锚越差这条线越高,它按平方往上跑。

    反过来,若锚是同源误差,全重叠给 1.00 倍,此时重叠是唯一能彻底解决的动作——加人到无穷也做不到零。两种情形下正确动作相反:同源谈共享测量流程,独立谈扩大校准样本。搞错方向,钱全白花。

    证据与局限

    已验证: 同源误差下膨胀等于根号下 1 加 K 乘 1 减 r,r 取 0、0.3、0.5、0.8、0.9、1.0 时实测 15.61、12.93、10.89、6.93、4.94、0.96,理论 15.52、13.00、11.00、7.00、5.00、1.00,偏差在 1% 内;独立误差式四组偏差在 3% 内;两式在 r 等于 0 处同给 15.5 倍,在 r 等于 1 处分别给 1.00 与 7.00;校准十人下 0 到 0.1 段降 0.8 倍、0.9 到 1.0 段降 4.0 倍;达两倍所需重叠率随校准人数单调下移,10/30/60/120/240 人对应 0.9875/0.9625/0.925/0.850/0.700;独立误差地板不随样本量下降(60/240/960 时 4.28/4.09/4.11)。

    部分验证: 全部计算建立在「真值共享、误差可加、锚为线性单参数」的线性高斯结构上,真实判读数据的非线性(切点漂移、有序分类)未纳入;「三分之二」是锚噪与读者噪之比为 0.25 平方时的值,比值一变这条线按平方移动,故是示例而非普适常数;重叠集合按随机抽取处理,若实际是「挑最容易的病例共享」(便利抽样),同样 r 下收益会低于随机重叠;「同源」与「独立」是两个极端,真实多为部分相关,中间态需相关系数加权,本篇未给连续族。

    待验证: ① 部分相关的锚误差(同设备不同批次、同师不同时段)下收益落在 1.00 与 7.00 之间何处,需一个可估的相关系数,而它在判读数据里从未被测过;② 多中心研究里「共享了多少校准病例」几乎从不报告,导致无法反查本篇曲线是否已被无意识验证过;③ 重叠病例若同时被用作培训材料,训练效应与校准效应混在同一个 r 里、方向可能相反,本篇未能分离。

    相关阅读

    证据等级: C
    参考文献: PMID 31732502, PMID 37972928, PMID 39285135, PMID 34111197, PMID 25627982
    L1 网络状态:CORRESPONDENCE_NETWORK.md(1931 行)逐词统计:「重叠」0 处、「交集」0 处、「泛化」0 处、「校准」1 处(第 1853 行 KB-687,指「共振频率是逐人个体量」的第一层含义,与校准子样本抽样无关)、「共享」2 处(第 1599 行指信息流对称共享、第 266 行指内感受量表多中心验证)、「多中心」2 处(第 266 行德国多中心 IAS 验证、第 979 行 964 例肝脾超声前瞻多中心),全部为研究设计描述而非误差传播层。该主题最接近的边仍是 E-060 标签前提(五脏辨证作分组标签|前提|标签信度,状态 conflict 未裁决)——KB-317 记两处错配(样本量单位错配、精度区间错配)、KB-334 要求一致性系数与患病率分报、PMID 25090041 指低基率下 κ 被严重惩罚;该边只登记「标签信度够不够用」,未登记「校准子样本之间的重叠如何改变误差传播、以及重叠共享的是真值还是测量事件」。EDGE_TABLE_science-yixue.md(63 行)grep「校准/标定/信度/一致/kappa/重测/方差/误差」全部 0 命中,概念最高止于生理层(频率 407 篇/谐波 147 篇/脉象 73 篇)。该主题尚未进入 L1 对应网络,已登记至 science-yixue-l1-gaps.md(「测量一致性」簇第 59 条)。


    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。

    Shared calibration cases buy not precision but \”the same truth\”: nine-tenths of the curve’s benefits lie in those final few percentiles.

    Evidence level: C

    The previous post offered an inexpensive prescription: instead of each center recruiting its own calibration cases, share a portion. This sounds reasonable, but no one has calculated the intermediate state. Once calculated, it collapses: sharing 30% of calibration cases buys almost nothing in the face of a fifteen-fold inflation. Moreover, what you think you are sharing are cases, but actually you are sharing only that person’s truth value—the error remains distinct for each.

    Main Body

    The conclusion of the previous post (#1422) was: calibrating ten people, each drawing independently, incurs a fifteen-fold cost for the precision difference between two readers, with the main factor being “how many people to calibrate.” It left a cost-saving exit—multi-center sharing of some calibration cases. This is algebraically correct, but the shape of the intermediate state has not been calculated.

    I. In the Square of Inflation, Overlap Only Buys a Linear Discount

    Two readers each take $n_{cal}$ calibration cases, with an intersection of $m$ cases; the overlap rate $r$ equals $m$ divided by $n_{cal}$. Two sources of error must be separated.

    If anchor error is the same for overlapping cases (same device, same senior reader viewing the same set of images), it is completely canceled in the subtraction of the two readers’ position estimates, yielding the closed-form solution: inflation equals the square root of $1$ plus $K$ times $1$ minus $r$, where $K$ equals (repetition count $\times$ sample size $\times$ $1$ plus anchor noise squared ratio) divided by (calibration count $\times$ reader noise squared).

    If anchor error is independent each time (same batch of cases, retested on different devices), overlap only cancels the truth part, leaving the anchor term unchanged: inflation equals the square root of $1$ plus $B$ times $1$ minus $r$ plus $A$, where $A$ is the anchor term and $B$ is the truth term.

    Both formulas match Monte Carlo cell-by-cell, with deviations within 1%. The key point is the first term: overlap appears only in the $1-r$ multiplier, which multiplies a base of tens to hundreds. The discount is real, but the base is huge.

    II. With 30% Overlap, Only Two and a Half of Fifteen Are Filled

    Taking samples of 240, reader noise of 0.5, two repetitions per person, and anchor noise of 0.5: with ten calibrators, $K$ equals 240, and inflation without overlap is 15.5-fold.

    As overlap moves from 0 to 0.3, inflation drops from 15.5 to 13.0. Saving 2.5-fold is real, but the remaining 13-fold is also real. To drop below two-fold, $r$ must equal 0.9875—meaning nine out of ten calibrators must be the same batch, which equals not sharing at all.

    Increasing calibration to thirty, sixty, and one hundred twenty people sets thresholds at 0.9625, 0.925, and 0.850 respectively. Only when the calibration count approaches the full sample (240 people, $K$ equals 10) does the threshold drop to 0.700, making sharing 30% meaningful.

    “Sharing” is not a substitute for “adding people”; it is downstream of it. Talking about sharing when the calibration count is not reached is saving money in a story that has not yet begun.

    III. The Shape of the Curve Is Reversed: Savings Come at the End, Not the Beginning

    By linear intuition, overlap should “save more at the beginning and less later.” In reality, the opposite is true—the derivative of inflation with respect to $r$ has a square root in the denominator, so as $r$ approaches 1, the denominator shrinks and the slope steepens.

    In this ten-calibrator cell: overlap from 0 to 0.1 reduces inflation by 0.8-fold; from 0.9 to 1.0, it reduces by 4.0-fold. The last 10% buys five times more than the first 10%.

    The implication is stark: a compromise like “sharing 30%” falls exactly on the flattest part of the entire curve. To achieve real gain, you must almost fully share; if you cannot, spend money on increasing the calibration count instead.

    IV. Two Universes Are Identical at Zero Overlap

    The two formulas reveal themselves where it matters most: at $r=0$, both give the same number, 15.5-fold; at $r=1$, one gives 1.00-fold and the other 7.00-fold.

    This means: before doing overlap, you cannot tell from the data which universe you inhabit. The entire value of this intervention is the difference between these two universes at full overlap. Under common-mode error, full sharing achieves zero cost; under independent error, full sharing still leaves a 7-fold floor.

    This floor does not decrease with sample size: samples from 60 to 240 to 960 yield measured values of 4.28, 4.09, and 4.11 respectively. Adding people, adding readers, or adding cases is all ineffective—it is not sampling error.

    Therefore, “sharing calibration cases” must first answer the prior question: what exactly is being shared. Sharing one case shares only that person’s truth value; to make errors cancel, you must share the same measurement event—same device, same reader, same time window. Multi-center studies each using their own devices to measure the same people share truth values, not errors.

    This explains the old lab experience: systematic bias in lipidomics across platforms and labs is suppressed by “common reference samples,” and it works because that sample is repeatedly measured on each platform using the same procedure—sharing measurement events, not samples themselves.

    V. When Is Overlap a Path?

    For independent error sources, the residual at full overlap is the square root of $1+A$. To drop below two-fold, $A$ must not exceed 3, meaning the calibration count must be at least (anchor noise squared $\times$ repetition count $\times$ sample size) divided by (three times reader noise squared). Under these parameters, this is 160 people, or 67% of the full sample.

    Two-thirds is the watershed where overlap shifts from “better than nothing” to “replacing adding people.” The worse the anchor, the higher this line, rising quadratically.

    Conversely, if the anchor is common-mode error, full overlap gives 1.00-fold, making overlap the only action that can fully solve the problem—even adding to infinity cannot reach zero. The correct actions are opposite in the two cases: common-mode requires sharing measurement procedures, independent requires expanding calibration samples. Getting the direction wrong wastes all money.

    Evidence and Limitations

    Verified: Under common-mode error, inflation equals the square root of $1+K$ times $1-r$. Measured values for $r=0, 0.3, 0.5, 0.8, 0.9, 1.0$ are 15.61, 12.93, 10.89, 6.93, 4.94, and 0.96, compared to theoretical values of 15.52, 13.00, 11.00, 7.00, 5.00, and 1.00, with deviations within 1%. For independent error, the four-group deviation is within 3%. Both formulas give the same 15.5-fold at $r=0$, but diverge to 1.00 and 7.00 at $r=1$. With ten calibrators, the drop from 0 to 0.1 is 0.8-fold and from 0.9 to 1.0 is 4.0-fold. The required overlap rate to reach two-fold monotonically decreases with calibration count: 10/30/60/120/240 correspond to thresholds of 0.9875/0.9625/0.925/0.850/0.700. The independent-error floor does not decrease with sample size (4.28/4.09/4.11 at 60/240/960).

    Partially Verified: All calculations rest on a linear Gaussian structure assuming “truth sharing, additive error, and anchor as a single linear parameter.” Nonlinearities in real adjudication data (threshold drift, ordered categories) are not included. The “two-thirds” value corresponds to the square of the ratio of anchor noise to reader noise being 0.25; if this ratio changes, the line shifts quadratically, so this is an example rather than a universal constant. Overlap sets are treated as randomly sampled; if in practice the easiest cases are shared (convenience sampling), the benefit at the same $r$ would be lower than random overlap. “Common-mode” and “independent” represent extremes; real cases are partially correlated, requiring weighted correlation coefficients for intermediate states, which this paper does not provide.

    To Be Verified: ① For partially correlated anchor error (same device different batches, same teacher different times), where the benefit falls between 1.00 and 7.00, requires an estimable correlation coefficient that has never been measured in adjudication data; ② In multi-center studies, “how many calibration cases were shared” is almost never reported, preventing verification of whether this paper’s curve has been unconsciously validated; ③ If overlap cases are also used as training material, training effects and calibration effects mix within the same $r$, possibly in opposite directions, which this paper cannot separate.

    Related Reading

    Evidence Level: C
    References: PMID 31732502, PMID 37972928, PMID 39285135, PMID 34111197, PMID 25627982
    L1 Network Status: CORRESPONDENCE_NETWORK.md (1931 lines) word-count: “overlap” 0 times, “intersection” 0 times, “generalization” 0 times, “calibration” 1 time (line 1853 KB-687, referring to the first layer of meaning that “resonance frequency is per-person individual quantity,” unrelated to calibration subsample sampling), “sharing” 2 times (line 1599 refers to information-flow symmetric sharing, line 266 refers to interoception scale multi-center validation), “multi-center” 2 times (line 266 German multi-center IAS validation, line 979 964 liver-spleen ultrasound prospective multi-center), all describing research design rather than error propagation layers. The closest edge is the E-060 label premise (organ differentiation as grouping label | premise | label credibility, status conflict undecided)—KB-317 records two mismatches (sample size unit mismatch, precision interval mismatch), KB-334 requires reporting consistency coefficient and prevalence separately, PMID 25090041 indicates $\kappa$ is severely penalized under low base rate; this edge only registers “is label credibility sufficient,” not “how does overlap between calibration subsamples alter error propagation, and what is shared—truth values or measurement events.” EDGE_TABLE_science-yixue.md (63 lines) grepping “calibration/calibration/credibility/consistency/kappa/retest/variance/error” yields zero hits; concepts stop at the physiological layer (frequency 407 papers/harmonic 147 papers/pulse 73 papers). This topic has not entered the corresponding L1 network, registered to science-yixue-l1-gaps.md (cluster “measurement consistency” item 59).


    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「零分布更宽所以假阳性更高」的直觉,方向反了:掺进来的亚群把 R₆ 压到了 0.5

    Evidence level: C

    R₆ 量的不是「分布有没有形状」,而是「亚群落在第六谐波的哪个相角」——同一个亚群,峰距 3 小时时它是负贡献,峰距 8 小时时它能放大上百倍。

    正文

    一、结果和预期反着走

    上一轮报 47.1% 假阳性率,前提是零假设人群纯正态、σ 取 1.43 小时。真实人群不是:褪黑素起始分型研究在 553 例里识别出 43 例与 42 例「不规则/多峰」剖面(PMID 42392022),时型问卷研究在 857 人里测出 16% 双峰型(PMID 28426387)。把零分布换成「主群正态 + 15% 双峰亚群」,更宽的零分布会把假阳性推到多少?

    按题目参数跑参数 bootstrap(σ=1.30、掺 15%、峰距 3.0 小时):n=200 时 R₆ 中位 0.52、P(R₆>1.5) 仅 1%,同档纯正态是 1.15 与 25%;n=60 是 0.90 对 1.44、20% 对 47%。95 分位更极端:峰距 3 小时时 P95 比纯正态低 35% 到 48%,离题目判据的「高出 50%」差一个数量级。假阳性不是升高,是降到二十五分之一。

    但这不能收下——同一模型峰距挪到 8 小时,P95 变成正 430% 到 210 万%。「更宽」不是这件事的变量。

    二、真正的变量是第六谐波的相角

    R₆ 是「实测第六阶合向量 ÷ 由 σ̂ 预测的第六阶」。对两成分混合,第六阶合向量约等于主群贡献加亚群贡献 p·cos(6θ),关键就是 cos(6θ) 这一项。题目设的峰距 3.0 小时是两侧各偏 1.5 小时,单侧相位 22.5°,六倍后 135°,cos 负 0.707。亚群对主群形成负贡献,把合向量拉小;分母只由 σ̂ 决定、照样变大,比值被两头挤压,掉到 0.5 以下。同一批参数若把亚群相位改成完全随机(不关于主群对称),R₆ 中位反而飙升到 282、P(R₆>4) 百分之百——对称性一破,相消就消失,噪声全被读成信号。

    逐点核对:峰距 2 小时(六倍角 90°)0.82;3 小时 0.53 最低;6 小时(270°)3.61;8 小时(360° 同相)27.05。曲线是周期 4 小时(24÷6)的振荡,不单调。R₆ 不是在测「分布有没有形状」,而是在测「形状落在第六谐波的哪个相角」——前几轮关于宽度、峰数、偏斜的直觉,在这一阶全部让位给相角。

    三、比混合更致命的是 σ

    上面全部计算都接受 σ=1.30 到 1.43。这个数站不住:美国时间使用调查十二年数据 n=53,689,按周末睡眠中点算时型离散度,20 到 24 岁组 σ=2.36 到 2.65 小时,75 岁以上最低也有 1.57 到 1.66 小时(PMID 28636610)。1.43 比全人群最低值还低。

    第六阶预测值对 σ 是二次指数敏感:σ=1.30 时 12.4%,1.43 时 8.0%,1.90 时 1.16%,2.65 时 0.017%。分母塌到千分之一,比值就飞到天上:σ=1.90、n=60 的纯正态样本 R₆ 中位 7.85、P(R₆>4) 为 75%;σ=2.31 时中位 60.4、P(>4) 为 98%。上一轮推荐的「R₆>4 判有形状」,在真实离散度下对完全正态样本判阳率七成到九成八。

    两因素叠加而非抵消:σ=1.90 落在相消区中位 7.83、同相区 23.1。那个「掉到 0.5」的相消,只在 σ 停在 1.30 附近才看得见。

    四、下尾是另一种失明,且参数不可辨识

    相消区内 n=200、σ=1.30、峰距 3 小时,P(R₆<0.5) 是 47%。「R₆ 明显低于 1」最自然的读法是「比正态还正态、没有形状」——但亚群实实在在存在,只是第六谐波贡献为负。纯正态下 P(R₆<0.5) 只有 8.0%,所以相消区的下尾是高得多的信号,只是方向指反了。现在整个分析体系里没有任何一处把「R₆ 显著小于 1」当作形状的证据,这条通路完全是空的。

    更硬的是辨识问题:R₆ 只交出两量(一阶与六阶合向量),生成机制却有主群 σ、占比 p、峰距三个未知。两方程三未知,解集是一维流形。在占比与峰距网格上找「R₆ 中位落在 1.5 上下」的组合,n=200 时找到 19 组:0.05 配 6 小时、0.20 配 4 小时、0.25 配 0.5 小时。同一读数既可以是「5% 的人偏 6 小时」,也可以是「25% 的人几乎不偏只是更散」。这是结构问题,多招人不会让流形收窄。文献侧的两成分 von Mises 混合从不压成一个比值,有研究报「80% 同相、20% 滞后约 8 小时」并用 bootstrap 判一峰两峰(PMID 16504088)。

    五、能改的只有两处

    第一,σ 必须每批自己估、单独报。前两轮分档表全钉在 1.43,人群实测跨 1.57 到 2.65,换一个 σ 能差三个数量级。做法是报 σ̂ 点估计与置信区间,按区间积分给出零分布带宽。

    第二,别用单一阶判形状。相消是阶特有的:峰距 4 小时在第六阶相消,但在第三阶同相。报整条残差谱并看过零点,比只看第六阶多一个约束。圈形数据方法学也已明确:多峰情形下常用单峰检验功效普遍不足,多峰判决要么用专门方法、要么老实拟合混合模型并报参数(PMID 30100666)。

    证据与局限

    已验证: R₆ 零分布对 σ 高度敏感(σ 由 1.43 升到 1.90,n=60 中位由 1.43 升到 7.85);混合亚群在第六阶产生 cos(6θ) 型相消或同相,周期 4 小时;纯正态下 P(R₆<0.5) 在 n=200 为 8.0%。均为本组参数 bootstrap(每格 1.2 万到 2 万次),基线复现上一轮(46.8% 对 47.1%、95 分位 3.83 对 3.82)。

    部分验证: 人群 σ 的 1.57 到 2.65 小时取自年龄分层时型离散度,测量对象是周末睡眠中点,与褪黑素起始相位不等价,本文只当量级参考。

    待验证: 真实人群的亚群峰距分布完全不已知——本文扫描了 0.5 到 12 小时全区间,但哪个区间是真实落点没有一篇文献报过。这是决定相消区是否真实存在的那一格。

    相关阅读

    证据等级: C
    参考文献: PMID 42392022, PMID 28426387, PMID 28636610, PMID 16504088, PMID 28860118, PMID 30100666
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 全文检索「零分布」「假阳性」「圆形统计」「合向量」「相位分布」「混合分布」均 0 命中;「单峰/双峰」仅 2 处,均在 E-060 附近的时频心肺耦合亚型语境(PMID 39583103),与昼夜相位零分布无关;EDGE_TABLE_science-yixue.md 概念最高止于生理层(频率/HRV/脉象),无统计零分布层。故本文封顶 C。

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


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

    The intuition that \”a wider null distribution leads to higher false positives\” is backwards: the injected subpopulation pushed R₆ down to 0.5

    Evidence level: C

    R₆ does not measure “whether the distribution has shape”; it measures “which phase angle of the sixth harmonic the subpopulation falls on”—the same subpopulation contributes negatively when peak separation is 3 hours but can amplify by over a hundredfold when peak separation is 8 hours.

    Body

    I. Results and expectations run in opposite directions

    The previous round reported a 47.1% false positive rate, assuming the null population is purely Gaussian with σ set to 1.43 hours. The real population is not: melatonin onset typology studies identified 43 and 42 cases of “irregular/multimodal” profiles among 553 subjects (PMID 42392022), and chronotype questionnaire studies detected 16% bimodal types among 857 individuals (PMID 28426387). Replacing the null distribution with “main group Gaussian + 15% bimodal subpopulation,” how far does the wider null distribution push the false positive rate?

    Running parameter bootstrap simulations per the problem parameters (σ=1.30, admixture 15%, peak separation 3.0 hours): at n=200, the median R₆ is 0.52, and P(R₆>1.5) is only 1%, compared to 1.15 and 25% for the same-tier purely Gaussian case; at n=60, it is 0.90 versus 1.44, and 20% versus 47%. The 95th percentile is even more extreme: with peak separation of 3 hours, the P95 is 35% to 48% lower than the purely Gaussian case, falling one order of magnitude short of the problem’s criterion of “50% higher.” False positives do not increase; they drop to one twenty-fifth.

    This cannot be accepted—when the same model’s peak separation shifts to 8 hours, the P95 becomes positive 430% to 2.1 million%. “Wider” is not the variable here.

    II. The true variable is the phase angle of the sixth harmonic

    R₆ is “measured sixth-order resultant vector ÷ predicted sixth-order based on σ̂.” For a two-component mixture, the sixth-order resultant vector approximates the main group contribution plus the subpopulation contribution p·cos(6θ), with the key term being cos(6θ). The problem sets peak separation at 3.0 hours, meaning each side deviates by 1.5 hours; the single-side phase is 22.5°, multiplied six times yields 135°, and cos(135°) is negative 0.707. The subpopulation exerts a negative contribution on the main group, pulling down the resultant vector; the denominator depends only on σ̂ and thus remains large, squeezing the ratio down below 0.5. If the same parameters shift the subpopulation phase to completely random (not symmetric about the main group), the median R₆ instead surges to 282, and P(R₆>4) reaches 100%—once symmetry breaks, cancellation disappears, and all noise is read as signal.

    Point-by-point verification: peak separation 2 hours (sixth-angle 90°) yields 0.82; 3 hours yields 0.53 (lowest); 6 hours (270°) yields 3.61; 8 hours (360°, same phase) yields 27.05. The curve oscillates with a period of 4 hours (24÷6), not monotonically. R₆ is not measuring “whether the distribution has shape”; it is measuring “which phase angle of the sixth harmonic the shape falls on”—earlier intuitions about width, peak count, and skewness all yield to phase angle at this order.

    III. More lethal than mixing is σ

    All calculations above accept σ=1.30 to 1.43. This value is untenable: the American Time Use Survey’s twelve-year data (n=53,689), calculating chronotype dispersion via mid-sleep on weekends, shows σ=2.36 to 2.65 hours for the 20–24 age group and as low as 1.57 to 1.66 hours for those over 75 (PMID 28636610). The value of 1.43 is lower than the lowest real population value.

    The sixth-order predicted value is quadratically exponential sensitive to σ: at σ=1.30 it is 12.4%, at 1.43 it is 8.0%, at 1.90 it is 1.16%, and at 2.65 it is 0.017%. The denominator collapses to one-thousandth, sending the ratio soaring: at σ=1.90 and n=60, the purely Gaussian sample’s median R₆ is 7.85, with P(R₆>4) at 75%; at σ=2.31, the median is 60.4, and P(>4) reaches 98%. The previously recommended “R₆>4 indicates shape” yields a false positive rate of 70% to 98% under real dispersion for completely Gaussian samples.

    The two factors compound rather than cancel: at σ=1.90, the cancellation zone median is 7.83 versus the same-phase zone at 23.1. That “drop to 0.5” cancellation is only visible when σ stays near 1.30.

    IV. The lower tail is another form of blindness, and parameters are unidentifiable

    In the cancellation zone with n=200, σ=1.30, and peak separation 3 hours, P(R₆<0.5) is 47%. The most natural reading of “R₆ significantly below 1” is “more Gaussian than normal, no shape”—but the subpopulation genuinely exists; it simply contributes negatively at the sixth harmonic. Under purely Gaussian conditions, P(R₆<0.5) is only 8.0%, so the lower tail of the cancellation zone is a much stronger signal, just pointing in the wrong direction. Nowhere in the entire analysis framework does “R₆ significantly less than 1” serve as evidence of shape; this pathway is entirely empty.

    More critically, there is an identification problem: R₆ yields only two quantities (first-order and sixth-order resultant vectors), yet the generation mechanism involves three unknowns: main group σ, proportion p, and peak separation. Two equations with three unknowns result in a one-dimensional manifold. Searching the proportion and peak separation grid for combinations where “the median R₆ falls around 1.5,” at n=200, we find 19 sets: 0.05 with 6 hours, 0.20 with 4 hours, 0.25 with 0.5 hours. The same reading can mean “5% of people deviate by 6 hours” or “25% of people hardly deviate but are simply more scattered.” This is a structural issue; increasing the number of recruits will not narrow the manifold. Literature on two-component von Mises mixtures never compresses into a single ratio; one study reports “80% same-phase, 20% lagging by about 8 hours” and uses bootstrap to distinguish unimodal from bimodal (PMID 16504088).

    V. Only two places can be changed

    First, σ must be estimated separately for each batch and reported individually. The first two rounds’ tiered tables are all pinned at 1.43, while real population values span 1.57 to 2.65; changing σ can differ by three orders of magnitude. The approach is to report the σ̂ point estimate and confidence interval, integrating over the interval to give the null distribution bandwidth.

    Second, do not use a single order to judge shape. Cancellation is order-specific: peak separation of 4 hours causes cancellation at the sixth order but is same-phase at the third order. Reporting the entire residual spectrum and checking zero crossings adds one more constraint compared to looking only at the sixth order. Circular data methodology has already clarified: under multimodal conditions, the power of unimodal tests is generally insufficient; multimodal judgment either requires specialized methods or honestly fits mixed models and reports parameters (PMID 30100666).

    Evidence and Limitations

    Verified: R₆ null distribution is highly sensitive to σ (σ rising from 1.43 to 1.90, n=60 median rises from 1.43 to 7.85); admixed subpopulations produce cos(6θ)-type cancellation or same-phase effects at the sixth order, with a 4-hour period; under purely Gaussian conditions, P(R₆<0.5) at n=200 is 8.0%. All via this group’s parameter bootstrap (12,000 to 20,000 per cell), baseline replicating the previous round (46.8% vs 47.1%, 95th percentile 3.83 vs 3.82).

    Partially verified: The population σ range of 1.57 to 2.65 hours comes from age-stratified chronotype dispersion, where the measurement target is mid-sleep on weekends, not equivalent to melatonin onset phase; this article uses it only as an order-of-magnitude reference.

    To be verified: The true population subpopulation peak separation distribution is completely unknown—this article scanned the full range from 0.5 to 12 hours, but no literature reports which interval is the true landing spot. This is the cell that determines whether the cancellation zone truly exists.

    Further Reading

    Evidence level: C
    References: PMID 42392022, PMID 28426387, PMID 28636610, PMID 16504088, PMID 28860118, PMID 30100666
    L1 network status: This topic has not yet entered the L1 corresponding network—full text search of CORRESPONDENCE_NETWORK.md for “null distribution,” “false positive,” “circular statistics,” “resultant vector,” “phase distribution,” “mixed distribution” yields zero hits; “unimodal/bimodal” appears only twice, both in the context of time-frequency cardiopulmonary coupling subtypes near E-060 (PMID 39583103), unrelated to the circadian phase null distribution; EDGE_TABLE_science-yixue.md concepts stop at the physiological layer (frequency/HRV/pulse), lacking a statistical null distribution layer. Hence this article caps at C.

    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.

  • 那个把 1.44 拉回 1.00 的校正,顺手把有形状的人也拉回了 1.00:这笔账从来没人定价

    Evidence level: C

    偏差校正修的是平均位置,付的是个体分辨率的账。这笔账在零分布上完全看不见,只有把真有形状的样本放进去才现形。

    正文

    一、先把母文的观察钉死

    上一篇实测:n=60 时正态族 R₆ 中位 1.44,基本 bootstrap 偏差校正(2R̂ 减重抽样均值)把它拉回 0.95 到 1.05,真值 1.0,修准了。代价写在括号里:校正后四分位距 0.29 到 1.65,比未校正的 0.95 到 2.14 更宽。

    我复现了这个数:n=60 未校正中位 1.492、校正后 0.990;n=200 是 1.234 到 1.085;n=480 是 1.024 到 0.966。零分布这一侧几乎完美。

    但「更宽」这句话不完整。我逐档量了标准差:n=200 时 0.489 到 0.581,涨 18.7%;n=480 涨 11.1%;n=120 涨 23.3%。最该出事的 n=30 反而暴涨 209.6%,从 2.069 到 6.406——同一档 24.8% 的校正值成了负数,而 R₆ 的物理下限是 0。

    二、零分布上看不出代价,因为代价根本不在这里

    只盯零分布会得出「校正几乎免费」的错觉。真正的代价在另一侧:把真有形状的样本放进去。

    n=60、真 R₆ 等于 3.48 的族(85% 主群加 15% 亚群、峰距 3 小时):未校正中位 5.74,校正后 1.51,压缩 74%。同构造 n=200 是 4.06 到 3.04;n=30 直接变负(负 0.90)。

    机制不难看出。校正量里的重抽样均值的期望不是常数——R 的分母是 S₁ 的 36 次方,而 S₁ 永远小于 1。重抽样时 S₁ 略降,S₁ 的 36 次方就塌一个量级,R 的副本随之爆大,均值被抬高,校正量过冲。对真值等于 1 的零分布,这个过冲恰好抵掉原有的正偏;对真值远大于 1 的形状族,同一个过冲把信号一起减掉。

    三、判别力掉多少:三条判据全部触线

    把「谁有形状」当分类问题,用对数 R₆ 做分数,算形状族对正态族的 AUC。

    n=60 时,强形状族(真 R₆ 12.2)AUC 从 0.991 掉到 0.835,降 0.156;中等族(3.48)从 0.870 掉到 0.738,降 0.132;峰距 6 小时的族(5.40)从 0.894 掉到 0.835。到 n=480,各族损失都回到 0.02 以内。

    检出力更难看。以各档校正后的 95 分位为阈:中等族 n=60 从 67.2% 掉到 31.6%,腰斩;n=200 从 85.6% 掉到 66.0%;n=480 才回到 86.4%。强形状族 n=60 从 96.8% 掉到 57.2%。要 80% 检出力,未校正时 n=120 够用(79.2%),校正后得 480。

    四、三条判据怎么读

    按预设的三条线逐条对:

    AUC 降超 0.15——n=60 强形状族 0.156,触线。偏差校正不能作单例判别的前处理,只能用于群体均值报告。

    兑换比跨族差异超 3 倍——实测从正 0.024 到负 924.8,跨四个数量级且变号。不存在统一校正策略,必须按族分别定价。

    所需 n 增超 2 倍——中等族从 120 到 480,四倍。「校正」在实践上等于「换更大样本」,对大样本无意义,只在小样本报告上有修辞价值。

    三条同时成立。校正把中位数修对了,代价是把个体分辨率先赔掉。它适合回答「这批人的平均 R₆ 是不是 1」,不适合回答「这个人有没有形状」。

    五、文献侧:这不是 R₆ 独有的病

    三篇方法学文献把这个结构讲得很清楚,只是没人把它搬到比值型圆形统计量上。

    第一篇(PMID 34302372)指出:主流做法是「点估计用 bootstrap 乐观校正、区间还用未校正的 DeLong 法」,两者口径不一致,覆盖率系统性低于名义水平,作者要求区间也做同样的校正。这句话翻过来就是——校正被当成单点数字的补丁,而不是整套推断的重写。

    第二篇(PMID 32401702)更狠:收缩类方法平均改善校准斜率,但样本间变异往往增大,且「估出来的收缩量」与「真正去过度拟合所需的最优收缩量」相关性通常是负的,作者结论是收缩不解决小样本问题。平均对了、个体更差,这个描述与本轮 R₆ 的压缩率完全同型。

    第三篇(PMID 24966219)给出边界:bootstrap、重复交叉验证、留对交叉验证在 EPV 十个以上或 C 统计量 0.9 以上时表现相近;更低时 bootstrap 偏乐观,但绝对误差最小。对应到本轮——只有当「信号强到 AUC 接近 1」时,校正的代价才可以忽略。 而脉象谐波这类研究的 AUC 恰在 0.85 附近,正落在代价最大的区间。

    六、落地

    不许拿校正值当单例判据。凡报校正后的 R₆,必须同时报四项:未校正值、校正后标准差、负值比例、以及本研究要回答的是群体均值还是个体判别。n 小于 120 直接禁报校正值——这一档里标准差涨幅 1.3%(n=60 零分布)与 209.6%(n=30)并存,说明校正量本身不稳,报出来只是一个更窄的数字带着更坏的分布。

    证据与局限

    已验证(本文自算,可复现;σ 取 1.43 小时,周期 24 小时,每格 250 次重复、每次 200 次重抽样):
    – 零分布侧校正有效:n=60 中位 1.492 到 0.990,n=200 是 1.234 到 1.085,n=480 是 1.024 到 0.966。
    – 零分布标准差升幅:n=30 涨 209.6%、n=60 涨 1.3%、n=120 涨 23.3%、n=200 涨 18.7%、n=480 涨 11.1%。
    – 校正后负值比例:n=30 达 24.8%、n=60 达 16.0%、n=120 达 8.8%、n=200 达 4.0%、n=480 达 0.4%。
    – 形状族被压缩:真 R₆ 3.48 的族 n=60 中位从 5.74 到 1.51(压缩 74%),n=30 变负。
    – AUC 损失(对数 R₆,形状对正态):n=60 强族降 0.156、中等族降 0.132;n=480 全部回到 0.02 内。
    – 检出力:中等族 n=60 从 67.2% 降到 31.6%、n=200 从 85.6% 降到 66.0%、n=480 为 86.4%。
    – 兑换比跨族从正 0.024 到负 924.8,符号翻转。

    部分验证:
    – 「乐观校正把点估计修好但不管区间」有方法学文献支撑(PMID 34302372):bootstrap 乐观校正的点估计被修,置信区间仍用未校正的 DeLong 法,覆盖率系统性低于名义水平。
    – 「收缩类方法平均改善、个体常更差」有模拟证据(PMID 32401702):校准斜率平均改善,但样本间变异常增大,且估计收缩与最优收缩的相关性通常为负,作者结论是收缩不解决小样本问题。
    – 「校正方法在低 EPV 或低 C 统计量时偏乐观」有实证(PMID 24966219):bootstrap、重复交叉验证、留对交叉验证在 EPV 十个以上或 C 大于 0.9 时表现相近,更低时 bootstrap 偏乐观但绝对误差最小。

    待验证:
    – 全部结论建在 σ=1.43 小时与特定族定义上;σ 换成实测的 1.57 到 2.65 小时后,R₆ 的绝对位置整体平移,压缩率与 AUC 损失需重算。
    – n=30 档的标准差暴涨是否真是重尾所致,我没做尾指数估计;也可能是 250 次重复不够。该档数字应视为「不稳定」而非「定量结论」。
    – 只测了基本 bootstrap 一种校正;BCa、.632+、留对交叉验证在此构造下的压缩率未知。PMID 37127679 指出留一交叉验证的 c 统计量可强烈偏零,方向与本轮一致但机制不同。
    – 「兑换比」我用了标准差升幅除以偏倚降幅,分母在某些族接近零,导致数值不稳;更稳的口径应是方差比的对数,本轮未做。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络 —— CORRESPONDENCE_NETWORK 全文「R₆」0 处、「合向量」0 处、「圆形统计」0 处、「相位分布」0 处、「昼夜相位」0 处、「重抽样」0 处;「判别力」与「AUC」各仅数处,全部落在脾弹性成像(PMID 35229196 / 36552916)与 E-060 的 HRV 复杂度语境,与偏差校正无关;EDGE_TABLE 概念最高止于生理层(频率 407 篇、谐波 147 篇、脉象 73 篇),无统计估计量层条目。按硬规则证据等级封顶 C。

    证据等级: C
    参考文献: PMID 34302372, PMID 24966219, PMID 32401702, PMID 41391734, PMID 37127679


    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 correction that pulled 1.44 back to 1.00 also pulled the shaped ones back to 1.00: this account was never priced

    Evidence level: C

    Bias correction fixes the average location, paying the bill for individual resolution. This bill is invisible on the null distribution; it only reveals itself when samples with true shape are included.

    Body

    I. First, nail down the observations from the parent text

    Previous measurement: with n=60, the median of the normal family R₆ was 1.44. Basic bootstrap bias correction (2R̂ minus resampled mean) pulled it back to 0.95–1.05, aligning with the true value of 1.0. The cost is written in parentheses: after correction, the interquartile range widened from 0.95–2.14 to 0.29–1.65.

    I reproduced these numbers: n=60 uncorrected median 1.492, corrected 0.990; n=200 ranged from 1.234 to 1.085; n=480 ranged from 1.024 to 0.966. The null distribution side is nearly perfect.

    But “wider” is incomplete. I measured standard deviations by tier: n=200 rose from 0.489 to 0.581 (+18.7%); n=480 rose +11.1%; n=120 rose +23.3%. The most problematic tier, n=30, exploded +209.6%, from 2.069 to 6.406—the same tier’s 24.8% correction value became negative, while R₆’s physical lower bound is 0.

    II. The cost is invisible on the null distribution because the cost lies elsewhere

    Focusing only on the null distribution creates the illusion that “correction is almost free.” The true cost lies on the other side: including samples with true shape.

    With n=60 and true R₆=3.48 (85% main group plus 15% subgroup, peak separation 3 hours): uncorrected median was 5.74, corrected median 1.51—a compression of 74%. Same construction with n=200 yielded 4.06 to 3.04; n=30 went directly negative (−0.90).

    The mechanism is clear. The expected mean of the resampled values in the correction amount is not constant—R’s denominator is S₁ raised to the 36th power, while S₁ is always less than 1. During resampling, S₁ decreases slightly, causing S₁³⁶ to collapse by an order of magnitude, inflating R’s copies and elevating the mean, leading to overcorrection. For the null distribution where true value equals 1, this overcorrection恰好 cancels the original positive bias; for shape families where true value is far greater than 1, the same overcorrection subtracts the signal along with the bias.

    III. How much did discriminative power drop? Three criteria all hit the line

    Treating “who has shape” as a classification problem, using log R₆ as the score, and calculating the AUC of shape families against normal families.

    With n=60, the strong shape family (true R₆=12.2) saw AUC drop from 0.991 to 0.835 (−0.156); the medium family (3.48) dropped from 0.870 to 0.738 (−0.132); the peak-separation-6-hour family (5.40) dropped from 0.894 to 0.835. By n=480, losses across all families returned within 0.02.

    Sensitivity looks worse. Using the post-corrected 95th percentile as threshold: medium family n=60 dropped from 67.2% to 31.6% (halved); n=200 dropped from 85.6% to 66.0%; only at n=480 did it return to 86.4%. Strong shape family n=60 dropped from 96.8% to 57.2%. To achieve 80% sensitivity, the uncorrected n=120 sufficed (79.2%), but after correction, n=480 is required.

    IV. How to read the three criteria

    Matching against the preset three lines:

    • AUC drop exceeding 0.15—the strong shape family at n=60 dropped 0.156, hitting the line. Bias correction cannot serve as pre-processing for single-case discrimination; it can only be used for reporting population means.
    • Exchange ratio cross-family difference exceeding 3×—measured from positive 0.024 to negative 924.8, spanning four orders of magnitude and changing sign. No unified correction strategy exists; pricing must be done family by family.
    • Required n increase exceeding 2×—medium family rose from 120 to 480 (fourfold). “Correction” in practice equals “switching to larger sample size,” rendering it meaningless for large samples and offering only rhetorical value for small-sample reporting.

    All three hold simultaneously. Correction fixes the median, but pays the bill by sacrificing individual resolution. It answers “Is the average R₆ of this group equal to 1?” but not “Does this individual have shape?”

    V. Literature side: this is not a disease unique to R₆

    Three methodological papers clarify this structure, though none have applied it to ratio-type circular statistics.

    The first (PMID 34302372) states: mainstream practice uses “bootstrap optimistic correction for point estimates, but uncorrected DeLong method for intervals,” creating inconsistent scales and systematic coverage below nominal levels; authors request interval correction as well. This inverted means—correction is treated as a patch for single-point numbers, not a rewrite of the entire inference framework.

    The second (PMID 32401702) is harsher: shrinkage methods average improve calibration slope, but inter-sample variability often increases, and the correlation between “estimated shrinkage amount” and “optimal shrinkage needed to avoid overfitting” is usually negative; authors conclude shrinkage does not solve small-sample problems. Average correct, individuals worse—this description matches the R₆ compression rate exactly.

    The third (PMID 24966219) sets boundaries: bootstrap, repeated cross-validation, and leave-one-out cross-validation perform similarly when EPV exceeds 10 or C-statistic exceeds 0.9; below that, bootstrap becomes optimistically biased but has minimal absolute error. Corresponding to this round—correction costs can be ignored only when “signal is strong enough that AUC approaches 1.” Since pulse wave harmonics research has AUC around 0.85, it falls precisely in the highest-cost zone.

    VI. Implementation

    Do not use correction values as single-case criteria. When reporting corrected R₆, simultaneously report four items: uncorrected value, post-corrected standard deviation, proportion of negative values, and whether the study aims to answer population mean or individual discrimination. Directly prohibit reporting correction values for n<120—this tier exhibits coexisting standard deviation increases of 1.3% (n=60 null distribution) and 209.6% (n=30), indicating the correction amount itself is unstable; reporting it yields only a narrower number with worse distribution.

    Evidence and Limitations

    Verified (self-calculated in this paper, reproducible; σ=1.43 hours, cycle 24 hours, 250 repetitions per tier, 200 resamples each):
    – Null distribution side correction effective: n=60 median 1.492 to 0.990, n=200 ranged 1.234 to 1.085, n=480 ranged 1.024 to 0.966.
    – Null distribution standard deviation increase: n=30 +209.6%, n=60 +1.3%, n=120 +23.3%, n=200 +18.7%, n=480 +11.1%.
    – Post-corrected negative proportion: n=30 reached 24.8%, n=60 reached 16.0%, n=120 reached 8.8%, n=200 reached 4.0%, n=480 reached 0.4%.
    – Shape families compressed: true R₆=3.48 family at n=60 median dropped from 5.74 to 1.51 (74% compression), n=30 became negative.
    – AUC loss (log R₆, shape vs normal): n=60 strong family −0.156, medium −0.132; n=480 all returned within 0.02.
    – Sensitivity: medium family n=60 dropped from 67.2% to 31.6%, n=200 from 85.6% to 66.0%, n=480 reached 86.4%.
    – Exchange ratio cross-family from positive 0.024 to negative 924.8, sign flipped.

    Partially verified:
    – “Optimistic correction fixes point estimates but ignores intervals” supported by methodological literature (PMID 34302372): bootstrap optimistic correction fixes point estimates, but confidence intervals still use uncorrected DeLong method, resulting in systematic coverage below nominal levels.
    – “Shrinkage methods average improve but individuals often worse” supported by simulation evidence (PMID 32401702): calibration slope averages improve, but inter-sample variability increases, and estimated shrinkage correlates negatively with optimal shrinkage; authors conclude shrinkage does not solve small-sample problems.
    – “Correction methods optimistically biased at low EPV or low C-statistic” supported by empirical evidence (PMID 24966219): bootstrap, repeated cross-validation, and leave-one-out cross-validation perform similarly when EPV>10 or C>0.9; below that, bootstrap is optimistically biased but has minimal absolute error.

    Pending verification:
    – All conclusions built on σ=1.43 hours and specific family definitions; replacing σ with measured 1.57–2.65 hours shifts absolute R₆ positions, requiring recalculation of compression rates and AUC losses.
    – Whether the standard deviation explosion at n=30 is truly due to heavy tails—I did not compute tail index estimates; it may also stem from insufficient 250 repetitions. This tier’s numbers should be treated as “unstable” rather than quantitative conclusions.
    – Only basic bootstrap correction was tested; BCa, .632+, and leave-one-out cross-validation compression rates under this construction remain unknown. PMID 37127679 notes that leave-one-out cross-validation c-statistic can be strongly biased toward zero, direction consistent with this round but mechanism different.
    – “Exchange ratio” used standard deviation increase divided by bias reduction; denominator approaches zero in some families, causing numerical instability; more stable metric should be log variance ratio, not computed in this round.

    Related Reading

    L1 Network Status: This topic has not yet entered the L1 corresponding network—CORRESPONDENCE_NETWORK lacks full entries for “R₆”, “sum vector”, “circular statistics”, “phase distribution”, “circadian phase”, “resampling”, “discriminative power”, and “AUC”; only scattered mentions exist, all falling within spleen elastography (PMID 35229196 / 36552916) and E-060 HRV complexity contexts, unrelated to bias correction; EDGE_TABLE concept reaches maximum at physiological layer (407 frequency papers, 147 harmonic papers, 73 pulse wave papers), lacking statistical estimator layer entries. By hard rules, evidence level capped at C.

    Evidence Level: C
    References: PMID 34302372, PMID 24966219, PMID 32401702, PMID 41391734, PMID 37127679


    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.

  • 那条直线的分子自带一个地板,分母的误差根本排不上号:零差异时「伤害」也有 0.29,而这已经比全部判据大六倍

    Evidence level: C

    上一轮担心分母算不准会把直线撑成喇叭口。算完发现:分母偏差不到百分之一,真正把这条线整块抬起来的是分子——两位读者完全一样,「伤害」也自动等于 0.29。

    一、先复述那个化简,再说它怕什么

    上一轮把判读一致性的天花板伤害压成一条近似直线:伤害约等于 TVD 除以 H。TVD 是两位读者边际的总变差距离,H 等于一减去合并边际各档比例的平方和。斜率为 1,于是五个百分点的容差可以直接换算:H 为 0.15 时,只容许 0.75 个百分点的 TVD。

    软肋很明显:H 不是常数,是从同批样本估出来的。分母若是估的,比值误差不等于两个误差各自除以自己,还要被协方差带着跑。上一轮留的正是这个问句——H 的误差有多大,会不会把直线撑成喇叭口。

    二、分母的偏差:方向对,量级错

    上一轮最担心的是凸性:样本版平方和系统性偏大,H 估偏小,伤害被高估。方向确实如此,量级却无关。三档、主档 0.92、各 71 例时,H 真值 0.1502,估计期望 0.1492,偏差 −0.69%;1000 例 −0.11%。近似式是负 H 除以总例数,一阶消失。

    上一轮设的线是「高估超过两成则全篇伤害数值整体下调」。−0.69% 差了两个数量级。这个方向虽然成立,却是整场核算里最不值得管的。

    三、分母的方差:才是它的软肋

    偏差小,方差大。H 的渐近方差有闭式:四倍的(三阶矩减二阶矩平方)除以总例数。代入三档情形,71 例标准误 0.0398、相对标准误 26.5%;200 例 15.8%;1000 例 7.1%。二十万次仿真给出 0.0398,与解析值逐位吻合。

    「相对误差超过一成则只能作定性排序」这条判据被触发了,且超标两点六倍。H 为 0.15 算出的容差换算,都带四分之一以上的不确定性。要压到一成以内需约五百例。这与 Simpson 多样性指数方差研究同向:该指数方差有闭式无偏估计,物种数超过样本量时朴素估计严重偏离(PMID 39020976)。

    四、真正的元凶在分子:它不归零

    把比值拆开就翻案了:TVD 在两位读者完全相同时,期望也不是零。两读者独立抽样,边际差不会恰好为零;单个档位上两比例绝对差的期望约等于根号下(4p 乘 1 减 p,再除以 n 乘圆周率)。各档相加折半,就是零差异地板。

    地板有闭式也有仿真:三档、主档 0.92、各 71 例,解析 0.0442,八万次仿真 0.0435。折算成伤害,地板除以 H 等于 0.29——而全部容差是五个百分点。两位一模一样的读者,光靠抽样噪声就造出 0.29,是判据的近六倍。加到 200 例降到 0.175,1000 例 0.078;要落进五百分点容差(TVD 0.0075),需约四千例。上一轮把 TVD 当作可无偏估的量、只怀疑分母;其实分子自带一个只按根号衰减的地板,在几十到几百例下比所有想判的效应都大。

    五、两者还纠缠,但方向与预期相反

    上一轮的分叉是:正相关则比值方差小于朴素合成,负相关则小 H 区剧烈放大。实测是正相关,近零差异时最强:无差异 +0.23 至 +0.35,真差异 0.04 时 +0.34,真差异 0.10 时 +0.46 至 +0.52,随样本量增大而减弱。

    正相关确实压掉一部分方差,可实际比值标准差仍是朴素合成的零点八到一倍,因为分子地板的加性抬升远大于这点压缩。问题不是「会不会放大」,而是总误差已被地板支配。名义 95% 区间对真值的覆盖:71 例时对称 delta 区间 88.7%、对数区间 79.2%、bootstrap 80.8%;200 例时 93.2%、84.7%、86.5%,三个都不达标。对数与 bootstrap 因地板非负更差——地板把分布压到零附近成一坨,假设对称或对数正态的区间都会失准。这与成对一致性度量区间的既有评估相符:低一致区 jackknife 反优于 bootstrap(PMID 21611165),层级二值结局的组内相关需非参数重抽样(PMID 16416402)。

    六、档数让地板更糟,不只是样本量

    档数对样本量的爆炸效应上一轮算过,地板这边有独立的坏消息:同一批材料、主档固定 0.92、余量均分,档数从 2 增到 6,地板从 0.0363 涨到 0.0602,折算伤害从 0.247 升到 0.395。

    原因是地板是各档项之和,每项按该档比例的根号计;余量摊成更多小档,项数增加快过单项缩小,且小比例项的相对波动更大。多档设计因此要付两笔账:上一轮那张样本量表算的是判据几何,这张算的是分子噪声,两笔同涨。档数对测量性质的影响早有提示:二到四档在多个心理测量指标上明显较差、七档后趋平(PMID 10769936)。

    实操上本篇给的不是「分母要加区间」,而是更前一道门禁:任何用伤害近似 TVD 除以 H 做五百分点级裁决的设计,先算零差异地板;地板超过判据一半,就不能判、只能描述。地板与样本量是平方根关系、与档数单调升,都不能靠事后解释绕开。

    证据与局限

    已验证:零差异下 TVD 地板有闭式解,三档 71 例解析 0.0442 与仿真 0.0435 吻合;H 的方差闭式与仿真逐位一致(71 例标准误 0.0398);H 偏差约等于负 H 除以总例数,71 例 −0.69%;比值区间名义 95% 在 71 例仅覆盖 79% 到 89%。

    部分验证:TVD 与 H 正相关(+0.23 至 +0.52),方向稳定、幅度依赖构型;档数 2 到 6 地板单调升(0.247 至 0.395),但只在余量均分这一构型上算过。

    待验证:真实材料的边际远比三档均分复杂,地板在偏斜、多峰、连续阈值切档下的形状未测;地板能否从材料清单事前预算,取决于能否拿到可信的边际先验;本篇全用独立抽样,若两读者共享参考答案或同一套培训,地板会因正相关而下降,降多少未算。

    相关阅读

    证据等级: C
    参考文献: PMID 8241374, PMID 39020976, PMID 21611165, PMID 16416402, PMID 10769936, PMID 8501467

    L1 网络状态:该主题尚未进入 L1 对应网络。逐词检索 CORRESPONDENCE_NETWORK.md(1163 行)与 EDGE_TABLE_science-yixue.md(63 行):TVD 0 处、总变差 0 处、异质 0 处、分母 0 处、抽样误差 0 处、simpson 0 处、kappa 0 处、边际 0 处;「一致」19 处全落在 E-060 标签前提(conflict,KB-292 判医师间一致性「低」且称令人担忧 PMID 30383398,对侧 KB-317/334 指出闸门规格与精度区间错配、AC2 在真实临床已跑通)与 E-061 频率共识两条边上。最接近的边为 E-060 标签前提(conflict),该边只登记「标签信度够不够用」,未登记「信度指标的抽样地板有多高」。本文不主张任何已验证机制,结论全部限定为方法学核算。


    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 straight line comes with a built-in floor in its numerator, so the denominator’s error is negligible: even at zero difference, the \”harm\” registers 0.29, which is already s…

    Evidence level: C

    Previous concern: denominator error might flare the straight line into a trumpet shape. Calculation shows: denominator error is less than 1/100th; the numerator actually lifts the entire line. Two readers with identical margins automatically yield “harm” = 0.29.

    I. Restate the simplification and its vulnerability

    The previous round compressed the ceiling of consistency-based harm into an approximate straight line: harm ≈ TVD / H. TVD is the total variation distance between readers’ margins; H equals one minus the sum of squared proportions across merged margins. The slope is 1, so a five-percentage-point tolerance converts directly: when H = 0.15, only 0.75 percentage points of TVD are allowed.

    The weakness is obvious: H is not constant but estimated from the same sample batch. If the denominator is estimated, the ratio error does not equal each error divided by itself; it is further dragged by covariance. The question left open previously was exactly this: how large is H’s error, and will it flare the line into a trumpet shape?

    II. Denominator bias: right direction, wrong magnitude

    The primary worry was convexity: the sample-based sum of squares systematically overestimates, making H underestimated and harm overestimated. The direction holds, but the magnitude is irrelevant. With three tiers, main tier 0.92, and 71 cases each, the true H is 0.1502, the estimated expectation is 0.1492 (bias −0.69%); with 1,000 cases it is −0.11%. The approximation formula is negative H divided by total sample size, vanishing at first order.

    The line set previously assumed “overestimation exceeding 20% would lower the entire harm scale.” A −0.69% difference spans two orders of magnitude. While this direction is valid, it is the least concerning part of the entire calculation.

    III. Denominator variance: the real weakness

    Bias is small; variance is large. H’s asymptotic variance has a closed form: four times (third moment minus squared second moment) divided by total sample size. Substituting the three-tier scenario, 71 cases yield standard error 0.0398 (relative standard error 26.5%); 200 cases 15.8%; 1,000 cases 7.1%. Twenty million simulations match the analytical value of 0.0398 digit-for-digit.

    The criterion “relative error exceeding 10% limits us to qualitative ranking” is triggered, and exceeded by 2.6 times. The tolerance conversion based on H = 0.15 carries over 25% uncertainty. To drop below 10%, approximately 500 cases are needed. This aligns with Simpson diversity index variance research: the index’s variance has a closed-form unbiased estimator, but when species count exceeds sample size, naive estimation deviates severely (PMID 39020976).

    IV. The true culprit lies in the numerator: it does not vanish to zero

    Breaking down the ratio overturns the verdict: TVD’s expectation is not zero even when two readers are identical. With independent sampling, marginal differences do not exactly vanish; the expected absolute difference of two proportions on a single tier is approximately √[(4p(1−p))/nπ]. Summing across tiers and halving gives the zero-difference floor.

    The floor has both closed-form and simulation values: three tiers, main tier 0.92, 71 cases each, analytical 0.0442, eight million simulations 0.0435. Converted to harm, floor divided by H equals 0.29—while the full tolerance is five percentage points. Two identical readers, relying solely on sampling noise, generate 0.29, nearly six times the criterion. At 200 cases it drops to 0.175, at 1,000 cases 0.078; to fall within the five-percentage-point tolerance (TVD 0.0075), approximately 4,000 cases are needed. The previous round treated TVD as an unbiasedly estimable quantity and doubted only the denominator; in fact, the numerator carries its own floor that decays only with the square root, dominating all intended effects at dozens to hundreds of cases.

    V. They still entangle, but direction and expectation reverse

    The previous split was: positive correlation yields ratio variance smaller than naive synthesis, negative correlation amplifies the low-H region. Measured reality shows positive correlation, strongest near zero difference: no difference +0.23 to +0.35, true difference 0.04 gives +0.34, true difference 0.10 gives +0.46 to +0.52, weakening as sample size grows.

    Positive correlation does suppress some variance, yet the actual ratio standard deviation remains 0.8 to one times the naive synthesis, because the additive lift from the numerator floor far outweighs this compression. The issue is not “whether it amplifies,” but that total error is already dominated by the floor. Nominal 95% intervals coverage of true values: 71 cases show symmetric delta interval 88.7%, log interval 79.2%, bootstrap 80.8%; 200 cases show 93.2%, 84.7%, 86.5%—none meet the target. Log and bootstrap intervals perform worse due to the non-negative floor, which compresses the distribution near zero into a lump; assuming symmetric or log-normal intervals becomes inaccurate. This matches existing assessments of paired consistency measure intervals: low-consistency regions see jackknife outperform bootstrap (PMID 21611165), and hierarchical binary outcome intraclass correlation requires nonparametric resampling (PMID 16416402).

    VI. Number of tiers worsens the floor, not just sample size

    The explosive effect of tier count on sample size was calculated previously; here the floor brings independent bad news: with the same material, main tier fixed at 0.92 and remainder evenly split, increasing tiers from 2 to 6 raises the floor from 0.0363 to 0.0602, converting harm from 0.247 to 0.395.

    The reason is that the floor sums across tiers, with each term scaled by the square root of its tier proportion; splitting the remainder into more small tiers increases the number of terms faster than individual terms shrink, and small-proportion terms exhibit larger relative fluctuation. Multi-tier design thus incurs two accounts: the previous round’s sample scale calculated the criterion geometry, this one calculates numerator noise; both rise together. The impact of tier count on measurement properties was hinted earlier: two to four tiers perform significantly worse across multiple psychometric indicators, seven tiers onward flatten (PMID 10769936).

    Practically, this paper does not add “denominator intervals,” but imposes an earlier gate: any design using harm approximated as TVD/H for five-percentage-point decisions must first compute the zero-difference floor; if the floor exceeds half the criterion, one cannot decide, only describe. The floor scales with the square root of sample size and monotonically increases with tier count; neither can be bypassed by post-hoc explanation.

    Evidence and Limitations

    Verified: The zero-difference TVD floor has a closed-form solution; three-tier 71 cases analytical 0.0442 matches simulation 0.0435; H’s variance closed-form matches simulation digit-for-digit (71 cases standard error 0.0398); H bias approximates negative H divided by total sample size (71 cases −0.69%); ratio intervals nominal 95% cover only 79% to 89% at 71 cases.

    Partially verified: TVD and H are positively correlated (+0.23 to +0.52), direction stable, magnitude depends on configuration; tier count 2 to 6 floor rises monotonically (0.247 to 0.395), but only for the evenly-split remainder configuration.

    To be verified: Real materials’ margins are far more complex than three-tier equal splitting; the floor’s shape under skew, multimodal, and continuous threshold-cut configurations remains unmeasured; whether the floor can be pre-budgeted from material lists depends on obtaining credible marginal priors; this paper uses independent sampling entirely—if two readers share reference answers or the same training, the floor would drop due to positive correlation, by how much remains uncalculated.

    Further Reading

    Evidence level: C
    References: PMID 8241374, PMID 39020976, PMID 21611165, PMID 16416402, PMID 10769936, PMID 8501467

    L1 network status: This topic has not yet entered the corresponding L1 network. Word-by-word search of CORRESPONDENCE_NETWORK.md (1,163 lines) and EDGE_TABLE_science-yixue.md (63 lines): TVD 0, total variation 0, heterogeneity 0, denominator 0, sampling error 0, simpson 0, kappa 0, margin 0; “consistency” 19 times all fall under E-060 label premise (conflict, KB-292 judges physician consistency “low” and calls it concerning PMID 30383398, contralateral KB-317/334 indicates gate specifications and precision intervals mismatch, AC2 already runs smoothly in real clinical settings) and E-061 frequency consensus edges. The closest edge is E-060 label premise (conflict), which registers only “is label reliability sufficient,” not “how high is the sampling floor of reliability metrics.” This paper does not advocate any verified mechanism; all conclusions are limited to methodological accounting.


    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.

  • 只要读者不是固定那两位,那条地板就整条作废:多读者轮换下「差」变成「散布」,而散布的零分布没人算过

    Evidence level: C

    「两位读者主档比例之差」这个量只在两人设计里存在。读者数一到三位以上,它有三个互不相同的替身,而这三个替身的零刻度能差到 2.65 倍——把两人设计的判据线搬过去,纯噪声会被判成「有稳定成分」,且读者越多判得越狠。

    一、两人差是个只在两人世界里成立的量

    前一篇把「δ 稳不稳」重写成「批次间方差 ÷ 含相关系数的二项方差」,零刻度定在 √(2−2ρ),判据线取 1.5 与 3.0。整套推导有一个被写死在开头的条件:读者恰好两位,且是固定的那两位。这个条件在真实判读研究里几乎不成立。三十位医师评三百多例病案的研究里,四百三十五对两两 κ 的中位只有 0.26,四分位区间 0.16 到 0.42,最小负 0.07、最大 0.62(PMID 22844445)——配对一多,配对本身就成了一种抽样。教育测量领域更直接:G 理论仿真给出的结论是,读者数少才是 ICC 偏倚与低效的主要来源,而不是题目数、也不是受试者数(PMID 33818118);而受试者数与读者数怎么分配,取决于读者方差与误差方差之比,比值不明时用相同数量最稳(PMID 16143966)。这些都是「加读者能提高精度」的证据,没有一条回答另一个问题:原来那条判据线,换了读者数还算不算数。

    二、三个替身的零刻度:降的、降得更快的、几乎不动的

    在等频三档、每批一百二十例、纯噪声的设定下,把三个替身的批间散布算到底(批次数放到四百,让估计收敛):两两配对取平均绝对差,读者数从二升到八,批间标准差从 3.88 个百分点降到 1.40,降了 64%;相对均值偏离(均方根)从 1.94 降到 1.11,本来就更窄、降幅也近似;边际极差从 3.88 只降到 3.72,几乎不动。极差与配对之比因此从 1.00 一路涨到 2.65。原题猜的是「极差期望随读者数上升、从二到八可涨近一倍」,实测不是近一倍,是 2.65 倍。原因是极差是极值型统计量:读者越多,最大最小两端各自往外走,多出来的读者主要贡献的是「更远的端点」,不是「更稳的中心」。这与一般化理论里读者方差被平均掉、误差被平均掉的直觉正相反——被平均掉的是中心,不是端点。

    三、把 1.5 搬过去,纯噪声就有九成八被判成有结构

    真正的灾难在搬运那一步。若沿用两人设计的地板去归一极差统计量,纯噪声下读者三位时已有 42.8% 越过 1.5 线;五位 89.0%;八位 97.8%,其中 22.2% 连 3.0 线都越了。也就是说,加读者本是为了让结论更可靠,实际效果是把「纯噪声」系统性地判成「有稳定成分」,而且读者越多判得越肯定。这个偏倚是单向的、有方向的、随设计改善而恶化的——最坏的一种错误。反过来,只要每个定义用它自己的零分布归一,三个定义在十二批次下越 1.5 线的概率都回到 1.1% 到 2.0%,名义水平恢复。所以问题不在多读者设计本身,而在判据线能不能跨定义搬。文献侧有同型证据:读者间异质性可以大到 eta 平方 0.40、超过被评者本身的 0.22,不做校正时假阴性率两成(PMID 40688984);评分者效应在多面设计里通常被报成「占比小于 5%」(PMID 26757931),但这个占比是方差占比,不是判据的零刻度。

    四、还有第二道闸:批次数太少,连正确的线都会被越

    归一化对了还不够。仍用配对定义、纯噪声,批次数四时批间散布的九十五分位是收敛值的一·六四倍,越 1.5 线的概率 8.5%;六批次 1.49 倍、4.9%;十二批次 1.35 倍、1.1%;二十四批次 1.24 倍、0.1%。也就是说「一点五」这条线在十二批次以下就已经被自身的抽样噪声吃掉一部分,四批次时几乎失效。这是一个原题没算的自由度:原题担心的是「读者数变了判据失效」,实测还叠了一层「批次少了判据自发越线」,两个方向都指向同一个结论——判据线不是常数,是(定义、读者数、批次数)三元组的函数。文献里读者数加到七位、G 系数从 0.86 升到 0.96,但五位已到 0.94,边际收益迅速递减(PMID 41181163),说明加读者不是免费的;而方差成分本身要六十位评分者、三十个刺激才估得稳(PMID 31898288),现实研究很少达到。

    五、三个定义不是三重证据,分工结论倒是跨读者数稳定

    最后一个反直觉点:配对、偏离、极差三个定义看起来像三条独立路径,实则共享同一批边际,高度相关。批乘读者交互漂移存在的条件下,三者「同时越线」的比例从读者两位的 2.4% 掉到八位的 0.5%——比单条路径还低,因为它们同向同源,一致通过不构成三重复现。另一方面,前一篇的分工结论跨读者数完全稳定:固定风格散布存在时,三个定义的越线率都停在噪声水平(1.2% 到 2.4%),散布型统计量对恒定风格偏移依旧零敏感;而位置型检验(批次均值之差相对其标准误)在同样条件下功效为 1.000。所以「散布管漂移、位置管偏移」这条分工不是两人设计的特例,换读者数照样成立;失效的只是那条被搬来搬去的判据线。

    证据与局限

    已验证:在阈值潜变量模型下,读者数由二增至八时,两两配对与相对均值偏离两种定义的纯噪声批间散布单调下降,边际极差几乎不动,两者之比由 1.00 升至 2.65;沿用两人设计地板归一极差统计量,读者八位时纯噪声越 1.5 线的概率达 97.8%;各定义用自身零分布归一后名义水平恢复至 1.1% 到 2.0%;批次数低于六时连正确归一化也会显著越线;固定风格偏移对三种散布型统计量零敏感,而位置型检验功效为 1.000。

    部分验证:文献支持「读者数少是 ICC 偏倚主因」(PMID 33818118)、「读者与受试者数量的最优配比依赖方差比」(PMID 16143966)、「读者间异质性可超过被评者」(PMID 40688984)、「读者数边际收益递减」(PMID 41181163),但没有任何一篇给出跨读者数的判据线换算表,也没有一篇报过极差统计量的零分布。

    待验证:一,本文的零分布来自等频三档加对称切点,非等频或偏斜切点下「读者数增大、极差零刻度上升」的斜率会不会变,未测;二,真实判读研究里读者是轮换而非全员评全批,稀疏连接下三种定义的方差还要再叠一层设计效应,本文按完全交叉算,属乐观估计;三,读者间存在系统分层(如资深与新手混编)时,极差会先被层间差撑大、再被层内波动填充,本文未分层。

    相关阅读

    证据等级: C
    参考文献: PMID 22844445, PMID 33818118, PMID 16143966, PMID 40688984, PMID 26757931, PMID 41181163, PMID 31898288

    L1 网络状态:CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md 中「读者」0 处、「rater」0 处、「判读」0 处、「极差」0 处、「轮换」0 处、「零分布」0 处、「kappa」0 处;「一致性」6 处、「信度」8 处全部落在 E-060 标签前提(五脏辨证作分组标签|前提|标签信度),该边状态为 conflict(未裁决):一侧为 PMID 30383398 判定开放人群中中医医师间诊断一致性「低(low)」、作者称「令人担忧」,另一侧为 KB-333/KB-334 指出「一致性 low」正落在加权系数精度最差区间、且 κ 型系数对序数/名义的敏感方向与 Gwet 一系相反,另有 PMID 31730403 等建议改用 Gwet AC2。该边只登记「标签信度够不够用」,未登记「换读者数后判据线的零刻度如何变」。据硬约束,本篇证据等级封顶 C,且不得写成已验证机制。

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


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

    The entire floor is invalidated unless readers are fixed to those two: under multi-reader rotation, \”difference\” becomes \”dispersion,\” and the zero distribution of dispersion ha…

    Evidence level: C

    The quantity “difference between the two readers’ main file proportions” exists only in designs with exactly two readers. When the number of readers reaches three or more, this quantity has three distinct proxies, and the zero scale of these three proxies can differ by up to 2.65 times—transplanting the criterion line from the two-reader design would cause pure noise to be judged as “having stable components,” and the more readers, the harsher the judgment.

    I. The Two-Reader Difference is a Quantity That Holds Only in the Two-Reader World

    The previous article rewrote “whether δ is stable” as “batch variance ÷ binomial variance containing correlation coefficients,” setting the zero scale at √(2−2ρ) and the criterion line at 1.5 and 3.0. The entire derivation rests on one condition hard-coded at the start: exactly two readers, and they must be fixed. This condition is almost never met in real-world judgment studies. In a study of 30 physicians rating over 300 cases, the median κ for 435 pairs was only 0.26, with interquartile range 0.16 to 0.42, minimum negative 0.07 and maximum 0.62 (PMID 22844445)—once pairing increases, the pair itself becomes a form of sampling. Educational measurement is even more direct: G-theory simulations conclude that fewer readers are the main source of ICC bias and inefficiency, not the number of items or subjects (PMID 33818118); and how to allocate subjects versus readers depends on the ratio of reader variance to error variance, so when this ratio is unknown, equal numbers are most stable (PMID 16143966). All these constitute evidence that “adding readers improves precision,” but none answers the other question: once the criterion line is transplanted, does it still hold?

    II. Three Proxies with Zero Scales: One Drops, One Drops Faster, One Almost Stays Still

    Under the equal-frequency three-tier, 120 cases per batch, pure noise setup, we computed the batch-to-batch spread of the three proxies to convergence (setting batches to 400): taking pairwise averages of absolute differences, as readers increase from two to eight, the batch standard deviation drops from 3.88 percentage points to 1.40—a 64% reduction; the relative mean deviation (root mean square) falls from 1.94 to 1.11, which was already narrower and the drop is approximately proportional; the marginal range shrinks only from 3.88 to 3.72, almost unchanged. Consequently, the ratio of range to pairwise average rises from 1.00 all the way to 2.65. The original conjecture was that “the expected range increases with reader count, nearly doubling from two to eight,” but the actual result is not nearly double—it is 2.65 times. This is because range is an extreme-value statistic: as readers increase, the maximum and minimum each move outward further; the extra readers mainly contribute “more distant endpoints,” not “more stable centers.” This contradicts the general theory intuition that reader variance gets averaged out and error gets averaged out—what gets averaged is the center, not the endpoints.

    III. Transplanting 1.5 Leads to 98% of Pure Noise Being Judged as Having Structure

    The real disaster occurs at the transplantation step. If we continue using the two-reader design floor to normalize the range statistic, under pure noise with three readers, 42.8% already exceed the 1.5 line; five readers: 89.0%; eight readers: 97.8%, of which 22.2% even cross the 3.0 line. In other words, adding readers was intended to make conclusions more reliable, but the actual effect is to systematically judge “pure noise” as “having stable components,” and the more readers, the more certain the judgment. This bias is unidirectional, directional, and worsens with design improvement—the worst kind of error. Conversely, if each definition uses its own zero distribution for normalization, then under twelve batches, the probability of exceeding the 1.5 line returns to 1.1% to 2.0%, restoring the nominal level. Thus the problem lies not in the multi-reader design itself, but whether the criterion line can be transplanted across definitions. Literature provides homologous evidence: reader heterogeneity can reach eta-squared 0.40, exceeding the subject’s own 0.22, and without correction, the false negative rate is 20% (PMID 40688984); rater effects in multifaceted designs are usually reported as “accounting for less than 5%” (PMID 26757931), but this proportion refers to variance contribution, not the zero scale of the criterion.

    IV. There Is a Second Gate: Too Few Batches, Even the Correct Line Gets Exceeded

    Normalization alone is insufficient. Still using the pairwise definition under pure noise, with four batches, the 95th percentile of batch-to-batch spread is 1.64 times the convergence value, giving an 8.5% probability of exceeding the 1.5 line; six batches: 1.49 times, 4.9%; twelve batches: 1.35 times, 1.1%; twenty-four batches: 1.24 times, 0.1%. In other words, the “1.5” line is already partially consumed by its own sampling noise below twelve batches, and almost completely ineffective at four batches. This is a degree of freedom not calculated in the original problem: the original worried about “reader count changes causing criterion failure,” but the actual result adds another layer—”fewer batches cause the criterion to spontaneously cross the line”—both directions pointing to the same conclusion: the criterion line is not constant, but a function of the triplet (definition, reader count, batch count). Literature shows that adding readers to seven increases the G coefficient from 0.86 to 0.96, but five already reaches 0.94, so marginal gains rapidly diminish (PMID 41181163), indicating that adding readers is not free; and variance components themselves require 60 raters and 30 stimuli to estimate stably (PMID 31898288), which is rarely achieved in real studies.

    V. The Three Definitions Are Not Triple Evidence, but Their Division of Labor Is Stable Across Reader Counts

    The final counterintuitive point: the pairwise, deviation, and range definitions appear to be three independent paths, yet they share the same marginal and are highly correlated. Under the condition that batch-by-reader interaction drift exists, the proportion of “simultaneous crossing” drops from 2.4% with two readers to 0.5% with eight—lower than any single path, because they originate from the same source and pass together does not constitute triple recurrence. On the other hand, the division-of-labor conclusion from the previous article remains completely stable across reader counts: when fixed style drift exists, the crossing rates of all three definitions stay at noise levels (1.2% to 2.4%), and the spread-type statistics remain zero-sensitive to constant style shifts; meanwhile, the location-type test (batch mean difference relative to its standard error) achieves power of 1.000 under the same conditions. Thus, “spread manages drift, location manages shift” is not a special case of the two-reader design—it holds regardless of reader count; what fails is only that transplanted criterion line.

    Evidence and Limitations

    Verified: Under the threshold latent variable model, as reader count increases from two to eight, the batch-to-batch spread of both pairwise and relative mean deviation definitions under pure noise monotonically decreases, while the marginal range remains almost unchanged; their ratio rises from 1.00 to 2.65; normalizing the range statistic using the two-reader design floor leads to 97.8% of pure noise exceeding the 1.5 line; after each definition is normalized using its own zero distribution, nominal levels recover to 1.1% to 2.0%; with fewer than six batches, even correct normalization significantly crosses the line; fixed style drift is zero-sensitive to all three spread-type statistics, while the location-type test achieves power of 1.000.

    Partially Verified: Literature supports “few readers are the main source of ICC bias” (PMID 33818118), “optimal ratio of readers to subjects depends on variance ratio” (PMID 16143966), “reader heterogeneity can exceed subject heterogeneity” (PMID 40688984), and “marginal gain of readers diminishes” (PMID 41181163), but no study provides a criterion-line conversion table across reader counts, nor any reports the zero distribution of the range statistic.

    To Be Verified: First, whether the slope of “increasing reader count raises the range zero scale” changes under unequal-frequency or asymmetric cutpoints, since this study uses equal-frequency three tiers with symmetric cutpoints; second, in real judgment studies readers rotate rather than all rate all batches, so under sparse connectivity the variance of the three definitions must also layer design effects, whereas this study assumes complete crossing and thus represents an optimistic estimate; third, when systematic stratification exists among readers (e.g., mixing experts and novices), the range is first inflated by inter-layer differences and then filled by intra-layer fluctuations, a scenario not addressed here.

    Related Reading

    Evidence Level: C
    References: PMID 22844445, PMID 33818118, PMID 16143966, PMID 40688984, PMID 26757931, PMID 41181163, PMID 31898288

    L1 Network Status: In CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md, the “reader” 0, “rater” 0, “judgment” 0, “range” 0, “rotation” 0, “zero distribution” 0, and “kappa” 0 entries; the “consistency” 6 and “reliability” 8 entries all fall under the E-060 label premise (organ differentiation as grouping label | premise | label reliability), and this edge status is conflict (unresolved): one side cites PMID 30383398, which judges diagnostic consistency among TCM physicians in the open population as “low (low),” with authors calling it “concerning”; the other side cites KB-333/KB-334, which indicates that “consistency low” falls precisely in the worst precision interval of weighted coefficient accuracy, and the ordinal/nominal sensitivity direction of κ-type coefficients opposes that of the Gwet series, with PMID 31730403 and others recommending switching to Gwet AC2. This edge only records “whether label reliability is sufficient,” but does not record “how the zero scale of the criterion line changes after switching reader counts.” Under hard constraints, this article’s evidence level caps at C and cannot be written as a verified mechanism.

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


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

  • 「错一格」和「错两格」被当成一样重:把档次当无序,等于把最有临床意义的那部分分歧抹成了同一个数

    Evidence level: C

    只要指标只读两个边际、不读表内结构,那么「轻度判成中度」与「轻度判成重度」对它贡献完全相同。这不是精度损失,是分辨力为零。

    一、一个被省掉的差别

    两位医生给同一批病人的严重程度打分,用的是「轻、中、重」三档。甲把某位本该「轻」的病人判成「中」,乙把同一位判成「重」。从病人的角度看,这两件事的重量差得很远:前者只是量级偏了一格,后者是方向性误判,可能直接改变处置路径。但在某类一致性指标里,这两个错误被记成同一个数,连小数点后第几位都不差。

    不是算错了,而是它读的数据不够。承 #1408 的代数:全部结论只用两个量,一是两读者边际的总变差距离 TVD,二是合并边际的辛普森异质性 H。这套写法换来一个漂亮性质——一致性上限只依赖这两个数,跟表里质量怎么摊开无关,数学上很干净。代价写在同一条定理里:既然上限只依赖边际,那么包括「错得远不远」在内的一切表内信息,都不在指标里。

    二、代数上这不是损失,是零

    先把这件事说死。设两位读者 A、B 的判读结果构成联合分布表,行和记作 r、列和记作 c,则 TVD 等于两者差额绝对值之和的一半,H 等于一减去合并边际 m(r 与 c 的平均)各项平方之和。两式只含 r 和 c,不含表内任何格子。

    逐位验证做过:固定同一对边际,让表内结构在十二种不同支撑集上跑——只在对角线放质量、只在错一格处放、只在跨两格处放、混合放,三到五档各一遍。TVD 小数点后六位一位不变,H 也一位不变。D=TVD 除以 H 这个量对表内结构的敏感度不是「很小」,是精确的零。

    这意味着「上限只依赖 D」在数学上没错,但它同时宣判了:只报 D 的报告,在比较两个判读系统时手里其实没有可用信息。因为两系统只要边际一样,D 就一模一样。

    三、留白到底有多宽

    那么在同一对边际下,一把有序尺子能在这个留白里走多远?可以精确算,不必靠抽样猜。把「行和等于 r、列和等于 c、每格非负」当约束,把加权 kappa 的分子写成权重点乘表内格子,就是一个线性规划:求它的最大最小值,得到的是解析界而非样本极值。

    算出来:三档、主档偏中(r 为 0.25、0.40、0.35,c 为 0.20、0.42、0.38)时,线性权 kappa 可达区间为负 0.4496 到正 0.9017,宽 1.3514;二次权为负 0.8182 到正 0.9301,宽 1.7483。四档宽 1.3924 与 1.8693,五档宽 1.3759 与 1.9381。三十万次随机抽样的实测极值与规划解吻合到小数点后三位,说明宽度是结构性的。

    一句话:边际相同、D 相同、上限相同,而有序一致性可在宽达一点四到一点九的区间里随意取。比任何一条判读门槛(0.6 或 0.7)都宽得多。

    四、最狠的一例:D 等于零

    如果两个读者的边际完全一致,TVD 等于零,D 也等于零。按无序读法这是最理想的情况:没有系统性偏移,判读分布严丝合缝。这时有序一致性被锁住了吗?没有。

    取主档 0.20、0.60、0.20 这组边际(两人分布相同),线性权 kappa 可达区间为负 0.25 到正 1.00,二次权为负 1.00 到正 1.00。两端都能取到:一端是两人都判对、只在中档微差错开,另一端是一人偏高一人偏低、错位互相抵消——后者在无序指标里完全隐形,在有序指标里是彻底失败。

    这是全文最该记住的一点,也比原题预设更强。原题假设「若存在成对的表使排序相反」,实测是凡边际相同的表对,无序指标必然打平,而有序指标的差可以任意大。不是偶发反例,是恒等成立的结构。

    五、排序反转接近抛硬币

    原题判据:排序相反比例超过一成五,则无序指标禁用于跨系统比较。实测远不止一成五。

    一千二百张合成表两两比较,剔除 D 打平的对(这类对无法排序,另行统计),可排序对的排序反转率:三档线性权 52.4%、二次权 52.7%;四档 49.3% 与 49.6%;五档 50.6% 与 50.4%。接近五成,跟抛硬币没区别。 一成五的线被超出三倍以上。

    更要紧的是打平对:四档出现一对、五档出现四对 D 完全相同的表,其有序指标差距中位数为 0.4879 与 0.2013,四档那一对百分之百超过 0.10。这正是原题担心的场景——「天花板相同」被读成「两系统同样好」。

    反向检验一下:两把有序尺子(线性权与二次权)会不会互相打架?会,但轻得多:方向相反率一成上下(10.5%、11.1%、10.9%),其中两边差距都超过 0.10 的只占 1.4%、0.8%、0.3%。矛盾主要不在「用哪把有序尺」,而在「要不要换有序尺」。

    六、判决与报告规范

    三条判据全部触发。

    第一,无序口径的 D 只可用于同一系统内的样本量规划,不得用于比较两个判读系统——排序反转率五成,比较结果与随机无异。第二,凡出现「两系统一致性上限相同」的表述,必须同时报有序加权指标,并以有序指标作决策依据。第三,若因样本太小只能报无序指标,须显式注明「档位按无序假设处理,跨格权重一律取一」。

    一个提醒:A 是错一格多、从不跨两格,B 是错一格少、但一错就跨两格。在 D 完全相同(0.05)的那组里,两者线性权 kappa 为 0.5156 与 0.5164,差不到千分之一;二次权却是 0.6610 与 0.3230,差 0.3380,远超 0.10 门槛。无序指标不是「稍微钝一点」,而是对最危险的那类分歧恰好失明。

    证据与局限

    已验证:TVD 与 H 仅由两个边际决定、与表内联合分布无关,这一点在固定边际、十二种支撑集、三到五档的设置下逐位成立;同一对边际下加权 kappa 的可达区间,用线性规划求得的精确端点与三十万次抽样实测极值吻合到小数点后三位;可排序对的排序反转率在三到五档、两种权重下均落在 49.3% 至 52.7% 之间。

    部分验证:线性权与二次权之间的方向相反率一成上下,其中双边差距均超 0.10 的仅 0.3% 至 1.4%,说明有序指标内部相对自洽;但这一结论依赖均匀 Dirichlet 生成的合成表分布,真实判读数据的边际偏斜与错位形态未必落在同一区域。

    待验证:一、加权 kappa 的可达区间能否写成只含边际与档数的解析式,若能则不必每次解规划;二、真实已发表的多读者有序一致性研究,其报告的线性权与二次权之差的实际分布落在可达区间的哪个位置,若普遍贴着一端,说明存在系统性的判读行为约束,可反过来当作设计检查。

    L1 网络状态:L1 网络中无「有序度量与无序度量的可辨识性」对应边。最接近的是 E-060 标签前提,其状态为「冲突(未裁决,且新增两处单位错配/精度区间错配)」,原文摘录:「KB-292(conflict):中医医师间诊断一致性在开放人群中『低(low)』……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」;该边只登记「标签信度够不够用」,未登记「指标把档位当无序会丢掉什么」。故本文证据等级封顶 C,上述全部结果为合成表上的代数与仿真结论,不构成对任何真实判读系统临床效力的判断。

    相关阅读

    证据等级: C
    参考文献: PMID 26095449, PMID 25168681, PMID 38283651


    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。

    \”Off by one\” and \”off by two\” are treated as equally heavy: treating tiers as unordered effectively erases the most clinically significant part of the divergence into a single n…

    Evidence level: C

    As long as the indicator reads only two margins and ignores the table’s internal structure, “mild misclassified as moderate” and “mild misclassified as severe” contribute identically to it. This is not a loss of precision; it is zero resolution.

    I. A Difference That Was Skipped

    Two doctors rate the severity of the same batch of patients using three tiers: “light,” “moderate,” and “heavy.” Doctor A misclassifies a patient who should be “light” as “moderate,” while Doctor B misclassifies the same patient as “heavy.” From the patient’s perspective, the weight of these two errors differs drastically: the former is merely an off-by-one magnitude error, whereas the latter is a directional misjudgment that may directly alter the treatment pathway. Yet in certain consistency indicators, both errors are recorded as the same number, differing not even in the decimal places.

    This is not a calculation error; it stems from insufficient data being read. Following the algebra of #1408, all conclusions rely on only two quantities: first, the Total Variation Distance (TVD) of the sum of the two readers’ margins, and second, Simpson’s heterogeneity (H) of the combined margins. This formulation yields a beautiful property—the upper bound of consistency depends solely on these two numbers, irrespective of how quality is distributed within the table, making it mathematically clean. The cost is stated in the same theorem: since the upper bound depends only on the margins, all internal table information—including “how far off” the error is—is excluded from the indicator.

    II. Algebraically, This Is Not Loss, It Is Zero

    Let us settle this once and for all. Suppose the joint distribution table of ratings by two readers, A and B, has row sums denoted as $r$ and column sums denoted as $c$. Then TVD equals half the sum of absolute differences between corresponding elements, and $H$ equals one minus the sum of squares of the combined margins $m$ (the average of $r$ and $c$). Both formulas depend only on $r$ and $c$, containing no cells from the table itself.

    Verification was performed bit by bit: fixing the same pair of margins, the internal table structure was run across twelve different support sets—placing quality only on the diagonal, only at the off-by-one position, only across two tiers, and mixed placements, covering three to five tiers each time. TVD remained invariant up to the sixth decimal place, and $H$ likewise remained constant. The sensitivity of the ratio $D = \text{TVD}/H$ to the internal table structure is not “small”; it is precisely zero.

    This means that while “the upper bound depends only on $D$” is mathematically correct, it simultaneously declares that any report reporting only $D$ lacks usable information when comparing two rating systems. Because both systems yield identical margins, $D$ is identical regardless of internal variation.

    III. How Wide Is the Blank Space?

    So, under the same pair of margins, how far can an ordered ruler traverse within this blank space? This can be calculated precisely without relying on sampling guesses. By treating “row sums equal to $r$, column sums equal to $c$, and every cell non-negative” as constraints, and writing the numerator of weighted kappa as the weighted sum of table cells, we obtain a linear programming problem: finding its maximum and minimum values yields analytical bounds rather than sample extremes.

    The results are as follows: with three tiers and the main tier biased toward moderate ($r = 0.25, 0.40, 0.35$; $c = 0.20, 0.42, 0.38$), the linear-weighted kappa ranges from $-0.4496$ to $+0.9017$, a width of $1.3514$; the quadratic-weighted kappa ranges from $-0.8182$ to $+0.9301$, a width of $1.7483$. For four tiers, widths are $1.3924$ and $1.8693$; for five tiers, widths are $1.3759$ and $1.9381$. Real-world extremes from 300,000 random samples matched the planned solutions up to the third decimal place, confirming that the width is structural.

    In one sentence: With identical margins and identical $D$, the upper bound is the same, yet ordered consistency can freely take any value within a range as wide as 1.4 to 1.9. This is far wider than any rating threshold (0.6 or 0.7).

    IV. The Most Extreme Case: D Equals Zero

    If the two readers’ margins are completely identical, TVD equals zero, and $D$ also equals zero. In terms of unordered reading, this is the ideal scenario: no systematic bias, and the rating distribution fits perfectly. Is ordered consistency locked in this case? No.

    Taking the main tier values of $0.20, 0.60, 0.20$ (where both readers have identical distributions), the linear-weighted kappa ranges from $-0.25$ to $+1.00$, while the quadratic-weighted kappa ranges from $-1.00$ to $+1.00$. Both extremes are achievable: one end corresponds to both readers agreeing, with only a slight off-by-one error in the middle tier; the other end corresponds to one reader being high and the other low, with errors canceling each other out—this latter case is completely invisible in unordered indicators but represents total failure in ordered indicators.

    This is the most important point of the entire text, and it is even stronger than originally assumed. The original hypothesis stated “if there exist pairs of tables with reversed ordering,” but the actual result is that for any pair of tables with identical margins, the unordered indicator must be tied, while the difference in ordered indicators can be arbitrarily large. This is not an occasional counterexample; it is a structural identity.

    V. Sorting Reversal Approaches Coin Tossing

    The original criterion stated: if the proportion of reversed sorting exceeds 15%, then the unordered indicator is unsuitable for cross-system comparison. In reality, this far exceeds 15%.

    Among 1,200 synthetic tables compared pairwise, excluding pairs where $D$ is tied (these cannot be sorted and are counted separately), the sorting reversal rate for sortable pairs is: 52.4% for linear weighting and 52.7% for quadratic weighting in three tiers; 49.3% and 49.6% in four tiers; and 50.6% and 50.4% in five tiers. This approaches 50%, indistinguishable from coin tossing. The 15% threshold is exceeded by more than three times.

    Even more critical are the tied pairs: in four tiers, one pair and in five tiers, four pairs had completely identical $D$. Their ordered indicator differences had medians of 0.4879 and 0.2013, with the four-tier pair exceeding 0.10 in all cases. This is precisely the scenario the original paper worried about—”same ceiling” being misread as “both systems equally good.”

    A reverse check: do the two ordered rulers (linear and quadratic weights) contradict each other? Yes, but lightly: the rate of opposite direction is around 10% (10.5%, 11.1%, 10.9%), with only 1.4%, 0.8%, and 0.3% having differences exceeding 0.10 on both sides. The main contradiction lies not in “which ordered ruler to use,” but in “whether to switch the ordered ruler.”

    VI. Judgment and Reporting Standards

    All three criteria are triggered.

    First, the $D$ of the unordered definition should be used only for sample size planning within the same system, not for comparing two rating systems—the sorting reversal rate is 50%, making comparison results indistinguishable from random. Second, whenever the statement “both systems have the same consistency upper bound” appears, the ordered weighted indicator must also be reported, and decisions must be based on the ordered indicator. Third, if the sample size is too small to report the ordered indicator, it must explicitly state “tiers are handled under the unordered assumption, and cross-tier weights are all set to one.”

    A reminder: A makes off-by-one errors but never crosses two tiers, while B makes fewer off-by-one errors but whenever it errs, it crosses two tiers. In the group where $D$ is completely identical (0.05), their linear-weighted kappas are 0.5156 and 0.5164, differing by less than one-thousandth; however, their quadratic-weighted kappas are 0.6610 and 0.3230, differing by 0.3380—far exceeding the 0.10 threshold. The unordered indicator is not “slightly dull”; it is precisely blind to the most dangerous type of disagreement.

    Evidence and Limitations

    Verified: TVD and $H$ depend solely on the two margins and are independent of the joint distribution within the table, confirmed under fixed margins, twelve support sets, and three to five tiers; the reachable interval of weighted kappa under the same margin pair, computed via linear programming, matches the extreme values from 300,000 random samples up to the third decimal place; the sorting reversal rate for sortable pairs falls between 49.3% and 52.7% across three to five tiers under both weighting schemes.

    Partially Verified: The rate of opposite direction between linear and quadratic weights is around 10%, with only 0.3% to 1.4% having bilateral differences exceeding 0.10, indicating internal self-consistency among ordered indicators; however, this conclusion relies on synthetic table distributions generated by uniform Dirichlet, and real rating data may exhibit margin skewness and misalignment patterns that do not fall within the same region.

    Pending Verification: First, whether the reachable interval of weighted kappa can be expressed as an analytical formula containing only margins and tier count, which would eliminate the need to solve linear programming repeatedly; second, for existing published studies on ordered consistency with multiple readers, where the actual distribution of differences between linear and quadratic weights falls within the reachable interval—if most cluster near one end, this suggests systematic constraints in rating behavior, which could serve as a design check.

    L1 Network Status: In the L1 network, there is no edge corresponding to “identifiability between ordered and unordered measures.” The closest is the E-060 label prerequisite, whose status is “conflict (unresolved, with two additional unit mismatches/precision interval mismatches).” The original excerpt states: “KB-292 (conflict): diagnostic consistency among TCM physicians in an open population is ‘low’… the second paper in the same series notes that chance-corrected consistency statistics are often misused and recommends Gwet AC2”; this edge only records “whether label reliability is sufficient,” not “what is lost when the indicator treats tiers as unordered.” Therefore, the evidence level of this paper caps at C; all above results are algebraic and simulation conclusions on synthetic tables, constituting no judgment on the clinical efficacy of any real rating system.

    Further Reading

    Evidence Level: C
    References: PMID 26095449, PMID 25168681, PMID 38283651


    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.

  • 补样本这一步到底废掉了什么:不是门槛,是「n=480」这句话

    Evidence level: C

    补样本没有把门槛绕过去。它绕过去的是「这是一份 480 例的研究」这句话——门槛没变,写进门的那张资格证变小了。

    一、先问对:废掉判据的到底是哪一半

    1405 给的验收办法是:不要用点估计 m̂≥0.95 就过线,要用 Clopper-Pearson 单侧精确下界。算下来零失败时 n=59 是硬下限,n=240 允许 6 例失败,n=480 允许 15 例。这套算术我复核过,全对。

    于是问题看起来很简单:实践中常见「先收 60 例,m̂ 只有 0.93,再补 60 例看看」,这个动作会让判据失效吗?

    要回答必须先拆开一件事:这套流程里有两个门槛同时在场,一个是要不要继续补的触发线,一个是最后放不放行的判定线。过去把它们当同一个东西,所以「补样污染判据」这个说法一直分不清是哪一个被污染。本轮把它们拆开做了模拟(八万次重抽,阶梯规则 60→120→240→480,触发线 m̂<0.95)。

    二、实测:判据对的那一半,补样几乎没污染它

    先看结论出人意料的一半。如果放行仍然用 #1405 的精确下界,只是允许中途补样,通过率是:真 m=0.90 时 0.2%,0.93 时 1.3%,0.95 时 4.7%,0.97 时 19.0%。

    拿它和固定 n=480 精确判据比:m=0.93 时固定是 0%,m=0.95 时 3.1%。补样版本的通过率不但没膨胀,还略高一点点,且始终压在名义 5% 附近。 换句话说,只要放行这一步用精确下界,中途补几次样并不会把 α 弄坏。backlog 预设的「通过率被抬到 60% 以上」不成立。

    原因不神秘:精确下界本身极难通过,补样加进来的样本也会被同一个下界一视同仁地审。污染发生在另一头。

    三、真正松的是点估计门禁,而它一直松

    换放行线为点估计 m̂≥0.95,同一套阶梯规则下:m=0.90 时 13.9%,m=0.93 时 44.5%,m=0.95 时 83.4%。

    看上去是灾难。但请再看固定样本那一列:n=60 时点估计门禁在真 m=0.93 处已有 38.9% 通过率;n=480 时降到 4.6%。同一个门禁,样本从 60 加到 480,误放率从 38.9% 掉到 4.6%——说明这个门禁的松紧几乎完全由 n 决定。

    所以补样不是「制造」了宽松,它是「保住」了宽松:一发现 m̂ 不够就补,等于始终在松的那一端做判决。

    四、补样的真实代价:等效样本量塌成四十例

    数据最刺眼的一处在这里。阶梯规则在真 m=0.93 处的通过率 44.5%,固定样本要到 n=40(46.0%)才匹配同样的宽松度;n=60 是 38.9%,n=120 只有 25.8%。

    也就是说这套流程名义上做到 480 例,实际检验力相当于 40 例。而报告里写的仍然是 480。

    代价立刻体现在两处。一是精度注水:通过者的平均实际 n 只有 82.1,按 82 例算标准误是 2.16 个百分点,按 480 例算只有 0.99——报告精度虚高 2.2 倍。二是取胜者诅咒:真 m=0.93 时,通过者平均报出的 m̂ 是 0.9609,比真值高 3.09 个百分点;同样条件下固定 480 例的通过者只高 0.71 个百分点,差 4.4 倍。把「补到过为止」的样本汇到一起报,得到的那个数系统性偏乐观。

    五、更硬的一条:精确下界本身在实践里不可用

    到这里才碰到真正的硬伤。#1405 的判据算术没错,但它没报功效。我算了一下:真 m 恰好等于门限 0.95 时,精确下界门禁的通过率恒为 α 本身——n=60 是 4.56%,n=240 是 4.22%,n=480 是 3.13%,n=960 是 3.98%。这是必然的:把门限当原假设,检验的显著性水平就是 α。

    要达到 80% 功效,真 m 得高到什么程度?n=59 需 0.99623,n=240 需 0.99907,n=480 需 0.99954。现实里不存在这种数据。 判读一致性研究的实际水平看,PMID 36123642 抽样 82 项一致性研究的样本中位数只有 71 例(Kappa 类)、仅 33% 给了样本量依据。

    结论很不好听:#1405 的门禁在数学上正确,在实践上等于「永不放行」。它不是验收工具,是证明工具——只能用来证明 m 大于 0.95,不能用来「验收 m 等于 0.95」。这两件事要的 α 和 n 完全不同。

    六、真序贯是唯一两头都站住的方案

    最后测了真序贯:补样触发和放行都用精确下界,一旦下界过线就停(放行),一旦上界已经低于 0.95 也停(无望,省样本)。结果 m=0.93 时通过率 1.6%,m=0.95 时 10.0%(平均 188 例),m=0.96 时 30.1%,m=0.97 时 70.0%(平均 268 例)。这才是有功效又有控制的形状——但注意它到 m=0.97 才敢放七成。

    所以本轮的裁决不是「禁不禁止补样」,而是先定门禁的性质,再谈流程。三条可落地的:

    第一,先声明这是验收还是证明。要「验收 m≥0.95」就必须接受 80% 以上放行率,那门限得从 0.95 下调(n=480 时降到约 0.921);要「证明 m>0.95」就写清楚它本就是保守判据,别拿来做通过率统计。

    第二,补样必须报「补了几次、每次的 n 和 m̂」,并且最终精度按实际 n 算,不按计划 n 算。这一条单独就能堵住 2.2 倍精度注水和 4.4 倍取胜者诅咒。

    第三,禁止「补到 m̂ 过线为止」。真要做,改用序贯上界(alpha spending,四次 look 用 0.005/0.010/0.020/0.050)。模拟显示它把 m=0.95 的通过率压回 4.3%,代价是功效也降——这是诚实的代价。

    证据与局限

    已验证:阶梯补样在精确门禁下通过率 1.3%(m=0.93)、4.7%(m=0.95),与固定 n=480 的 0%、3.1% 同量级,α 未被污染;点估计门禁在 n=60 时对 m=0.93 已有 38.9% 误放;门限处精确门禁通过率恒为 α(3.1%–4.6%);80% 功效所需真 m 为 0.996 以上。

    部分验证:等效样本量塌缩(阶梯 44.5% 对应固定 n≈40)依赖具体阶梯参数(60/120/240/480 与 0.90/0.93/0.95 三档阈值),换参数数值会变,但「等效 n 远小于计划 n」的定性结论在几组参数下都成立。

    待验证:真实研究里「补样」的比例无人统计——PMID 36123642 只报 33% 给了样本量依据,没说这 33% 是不是事前定的;取胜者诅咒在真实发表数据中的幅度也无人量过;「验收型门限下调到多少」需要按各领域可接受的误放率单独定,本轮的 0.921 只是按 80% 放行反解出来的一个示例值。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络。边表中最接近的一条是 E-060 标签前提,其原文为「冲突(未裁决,且新增两处单位错配/精度区间错配)」,只裁决标签整体信度与训练能否改善,从未涉及样本量确定方式与可选停止;对应网络全文检索样本量、序贯、停止规则、预注册等词均无命中。故本文证据等级封顶 C,结论属方法学推演而非已证机制。

    证据等级: C
    参考文献: PMID 24338600, PMID 36123642, PMID 15733050, PMID 39744919, PMID 34494220, PMID 28183155, PMID 23824994


    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。

    What exactly does this step of supplementing samples destroy: not the threshold, but the phrase \”n=480\”.

    Evidence level: C

    Augmenting the sample does not bypass the threshold. It bypasses the claim that “this is a study of 480 cases”—the threshold remains unchanged, but the certificate stamped at the door becomes smaller.

    I. First, Ask Correctly: Which Half of the Criterion Was Actually Abolished

    The acceptance criterion given in #1405 is: do not pass based on the point estimate $\hat{m} \geq 0.95$; instead, use the Clopper-Pearson one-sided exact lower bound. Calculating this yields a hard floor of $n=59$ for zero failures, $n=240$ allowing 6 failures, and $n=480$ allowing 15 failures. I have verified this arithmetic; it is entirely correct.

    Thus the apparent problem is simple: in practice, it is common to “first collect 60 cases, $\hat{m}$ is only 0.93, then augment another 60 cases to see.” Does this action invalidate the criterion?

    To answer, one must first disentangle a specific issue: within this procedure, two thresholds are simultaneously present—one is the trigger line for whether to continue augmenting, and the other is the decision line for whether to finally release. In the past, these were treated as the same thing; hence the phrase “sample augmentation pollutes the criterion” remained ambiguous about which part was polluted. This round separates them and performs simulations (80,000 re-samples, stepwise rules 60→120→240→480, trigger line $\hat{m} < 0.95$).

    II. Empirical Test: The Half of the Criterion That Was Correct Is Almost Unpolluted by Augmentation

    First, look at the half of the conclusion that is surprising. If release still uses #1405’s exact lower bound, merely allowing mid-study augmentation, the pass rates are: true $m=0.90$ yields 0.2%, $m=0.93$ yields 1.3%, $m=0.95$ yields 4.7%, and $m=0.97$ yields 19.0%.

    Comparing this to the fixed $n=480$ exact criterion: $m=0.93$ yields 0%, and $m=0.95$ yields 3.1%. The pass rate of the augmentation version does not inflate; it is slightly higher, and it remains pressed near the nominal 5%. In other words, as long as the release step uses the exact lower bound, mid-study augmentation several times will not ruin $\alpha$. The backlog-presumed “pass rate inflated above 60%” does not hold.

    The reason is not mysterious: the exact lower bound itself is extremely hard to pass; samples added via augmentation are also subjected to the same lower bound uniformly. Pollution occurs on the other end.

    III. What Truly Relaxes is the Point Estimate Gate, and It Has Always Been Loose

    Switching the release line to the point estimate $\hat{m} \geq 0.95$, under the same stepwise rules: $m=0.90$ yields 13.9%, $m=0.93$ yields 44.5%, and $m=0.95$ yields 83.4%.

    This looks like a disaster. But please look again at the fixed-sample column: at $n=60$, the point estimate gate already has a 38.9% pass rate at true $m=0.93$; at $n=480$ it drops to 4.6%. The same gate, with sample size increasing from 60 to 480, sees the error-release rate drop from 38.9% to 4.6%—indicating that the looseness of this gate is almost entirely determined by $n$.

    Therefore, augmentation does not “create” looseness; it “preserves” looseness: once $\hat{m}$ is found insufficient, augmentation equals always making the decision at the loose end.

    IV. The Real Cost of Augmentation: Equivalent Sample Size Collapses to Forty Cases

    The most striking data point lies here. Under the stepwise rule, at true $m=0.93$ the pass rate is 44.5%; to match this same looseness, the fixed sample must reach $n=40$ (46.0%); $n=60$ is 38.9%, and $n=120$ is only 25.8%.

    This means the procedure nominalistically reaches 480 cases, but its actual statistical power is equivalent to only 40 cases. Yet the report still states 480.

    The cost immediately manifests in two places. First, precision inflation: the average actual $n$ of passers is only 82.1; calculating standard error at 82 yields 2.16 percentage points, whereas at 480 it is only 0.99—the reported precision is inflated by 2.2 times. Second, the winner’s curse: at true $m=0.93$, passers report an average $\hat{m}$ of 0.9609, which is 3.09 percentage points higher than the true value; under the same conditions, fixed $n=480$ passers are only 0.71 percentage points higher, a difference of 4.4 times. Pooling samples “augmented until $\hat{m}$ crosses the line” yields a number that is systematically overly optimistic.

    V. A Harder Point: The Exact Lower Bound Itself Is Unusable in Practice

    Only here do we encounter the real hard flaw. The arithmetic of #1405’s criterion is correct, but it does not report power. I calculated: when true $m$ exactly equals the threshold 0.95, the pass rate of the exact lower-bound gate is constantly equal to $\alpha$ itself—$n=60$ yields 4.56%, $n=240$ yields 4.22%, $n=480$ yields 3.13%, and $n=960$ yields 3.98%. This is inevitable: treating the threshold as the null hypothesis means the test’s significance level is $\alpha$.

    To achieve 80% power, how high must true $m$ be? $n=59$ requires 0.99623, $n=240$ requires 0.99907, and $n=480$ requires 0.99954. Such data do not exist in reality. Looking at the actual level of consistency studies, PMID 36123642 reports that among 82 sampled consistency studies, the median sample size is only 71 (Kappa-type), and only 33% provided a sample-size basis.

    The conclusion is unflattering: #1405’s gate is mathematically correct but in practice equals “never release.” It is not an acceptance tool; it is a proof tool—used only to prove $m > 0.95$, not to “accept $m = 0.95$.” These two tasks require completely different $\alpha$ and $n$.

    VI. True Sequential Is the Only Solution That Holds Both Ends

    Finally, we tested true sequential: both augmentation trigger and release use the exact lower bound; stop once the lower bound crosses (release), or once the upper bound falls below 0.95 (hopeless, saves samples). Results: $m=0.93$ yields 1.6%, $m=0.95$ yields 10.0% (average 188 cases), $m=0.96$ yields 30.1%, and $m=0.97$ yields 70.0% (average 268 cases). This is the shape with both power and control—but note it only dares release 70% at $m=0.97$.

    Thus the verdict of this round is not “whether to ban augmentation,” but first to define the nature of the gate, then discuss the procedure. Three actionable points:

    First, declare upfront whether this is acceptance or proof. To “accept $m \geq 0.95$” one must accept a release rate above 80%; then the threshold must be lowered (at $n=480$ it drops to about 0.921). To “prove $m > 0.95$” one must clearly state it is inherently a conservative criterion and not use it for pass-rate statistics.

    Second, augmentation must report “how many times augmented, each time’s $n$ and $\hat{m}$”, and final precision must be calculated based on actual $n$, not planned $n$. This single rule alone can block the 2.2-fold precision inflation and the 4.4-fold winner’s curse.

    Third, prohibit “augment until $\hat{m}$ crosses the line”. If truly necessary, use sequential upper bounds (alpha spending; four looks use 0.005/0.010/0.020/0.050). Simulation shows this presses the $m=0.95$ pass rate back to 4.3%, at the cost of reduced power—this is the price of honesty.

    Evidence and Limitations

    Verified: Under exact gates, stepwise augmentation yields pass rates of 1.3% ($m=0.93$) and 4.7% ($m=0.95$), comparable to fixed $n=480$’s 0% and 3.1%; $\alpha$ is not polluted. The point estimate gate at $n=60$ already has 38.9% error-release for $m=0.93$. At the threshold, exact gate pass rate is constantly $\alpha$ (3.1%–4.6%). Achieving 80% power requires true $m > 0.996$.

    Partially Verified: Equivalent sample size collapse (stepwise 44.5% matching fixed $n \approx 40$) depends on specific step parameters (60/120/240/480 and thresholds 0.90/0.93/0.95); changing parameters alters values, but the qualitative conclusion that “equivalent $n$ is far less than planned $n$” holds across several parameter sets.

    Pending Verification: In real studies, the proportion of “augmentation” is unstated—PMID 36123642 reports only 33% provided sample-size basis, but does not say whether this 33% was pre-specified; the magnitude of the winner’s curse in real published data is also unmeasured. “How much to lower the acceptance threshold” requires individually setting based on acceptable error-release rates per field; the 0.921 in this round is merely an example derived from reversing 80% release.

    Further Reading

    L1 Network Status: This topic has not yet entered the L1 corresponding network. The closest entry in the edge table is the E-060 label premise, whose original text reads “conflict (unresolved, plus two new unit mismatches/precision interval mismatches),” which only adjudicates overall label credibility and whether training can improve, never involving sample-size determination methods or optional stopping; corresponding network full-text searches for “sample size,” “sequential,” “stopping rules,” “pre-registration,” etc., yield no hits. Hence this article’s evidence level is capped at C; its conclusions belong to methodological deduction rather than proven mechanisms.

    Evidence Level: C
    References: PMID 24338600, PMID 36123642, PMID 15733050, PMID 39744919, PMID 34494220, PMID 28183155, PMID 23824994


    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「可接受错档幅度」到底是不是一个数:答案取决于你把它记在哪把尺子上

    Evidence level: C

    同一个领域差异,记在「错档率」上放大 1.51 倍、记在「主档占比差」上只放大 1.12 倍——所以「阈值该不该分领域」这个问题,先要问的是记账口径,不是经验数据。

    正文

    一、问题的提法。 前一篇(#1411)的整个推导建立在一个从未被检验的假设上:存在一个统一的 w,即「可接受的错档幅度」,全领域共用,只是取值待定——0.05 还是 0.15。更麻烦的可能性是:w 根本不是一个数,而是一个随判读任务变化的函数。影像判读里错一档的代价是「多做一个检查」,功能评定里错一档的代价是「换一套康复方案」,两者临床代价差一个量级,凭什么用同一个 w?如果 w 真依情境而变,那原来的第二条处方(不许跨主档搬运边界)就要升级成更强的禁令:不许跨领域搬运,连 0.05 这个「通用值」都只是某个领域的历史产物被误当普适。

    这个怀疑不是空想。耳鼻喉文献里 35 篇最小临床重要差研究,计算与报告方式差异极大:82.9% 用锚定法、68.6% 用分布法,只有一半两者兼用,仅 25.7% 报出该阈值的敏感度与特异度(中位 60.8% 与 80.4%)(PMID 37933798)。同一个概念在不同研究里算出的数并不一致,且多数研究连它准不准都没验。更早的方法学综述已经把话说死:分布法多数实际算的是「最小可检测变化」而非「最小重要变化」,而后者依锚类型、最小重要性的定义、疾病特征而变(PMID 16925807)。

    二、第一次仿真:把 w 记在错档率上。 我按最直觉的口径做仿真:w 定义为「错档率」,即两读者档位不一致的病例占比。三个「领域」用判读噪声区分(0.45、0.75、1.05 个潜变量标准差,大致对应影像、功能评定、心理量表三类任务的难度差),每域两万批、每批六十例。

    结果:错档率从影像的 0.410 升到心理量表的 0.621,放大 1.51 倍。原题判据说「领域间 w 中位差超过 2 倍」才须分领域建表,1.51 倍擦着线但没到。不过 1.5 倍这个门槛本身也是随手定的,1.51 已足够说明 w 不是一个数。看起来怀疑成立。

    三、换个口径,结论塌了。 把 w 改成前一篇实际使用的口径——「两读者主档占比之差」(百分点)——重跑同一套数据。

    结果:占比差从影像的 3.02 个百分点只升到心理量表的 3.37 个百分点,放大 1.12 倍。同一个领域差异、同一个 1.5 倍判据,这次连边都够不着。

    更关键的是地板。批级占比差的中位数在三个领域都是 5.00 个百分点,一模一样。而它的理论地板由主档占比 p 与批样本量 n 决定:p 等于 0.45 时 9.08 个百分点,p 等于 0.15 时 6.52 个百分点。这把尺子的分辨率完全由「主档有多满、批有多大」决定,与判读任务难不难几乎无关。

    四、机制:为什么两把尺子给出相反答案。 因为占比差是减法。两位读者各自偏离真值,任务越难偏离越大;但「任务难」这个分量是两人共有的——同一批病例对谁都难——于是在相减时抵消掉了,留下来的只有两人之间的差异部分。而错档率是绝对误差,共有分量不但不抵消,还双倍进入。

    这解释了一个一直没人说破的事:为什么上一轮反复强调「答案藏在没人报的东西里」——已发表研究几乎只给一致性系数与区间,不给两读者各自的完整边际。因为一旦给了边际,占比差就能算;不给,就只能退回错档率类指标,而那类指标会把领域难度整个吃进去。于是「领域差异」在文献里看起来很大,很大一部分其实是记账方式造成的假象。

    五、判据要改写。 原题三条判据全部失效,因为它们的结论取决于 w 记在哪个量上。第一条「领域间 w 中位差超 2 倍则分领域建表」:错档率口径下 1.51 倍不触发,占比差口径下 1.12 倍更不触发——但原因完全不同,一个是擦线,一个是根本没反应。第二条「同一份数据两领域判出相反结论的比例超 30% 则跨域元分析不可信」:占比差口径下,w 相差 2 倍时矛盾率 0.256 至 0.273,不到 30%;相差 3 倍时 0.475 至 0.488,远超 30%。而错档率口径下,无论相差 1.5 倍还是 5 倍,矛盾率都在 0.86 至 0.98 之间,判据永远触发。同一个阈值,一个口径下是「偶尔冲突」,另一个口径下是「永远冲突」。

    所以真正的结论既不是「w 必须分领域」,也不是「w 可以通用」,而是:「w 该不该分领域」这个问题的答案由记账口径决定,不由经验数据决定。 你先声明 w 记在哪个统计量上,才轮得到问它跨不跨领域。这与上一轮的教训同型:坐标选错,全部结论反向。

    证据与局限

    已验证: 上述倍数关系与地板公式来自本轮仿真(每域两万批、每批六十例,四档有序量表、潜变量模型),算术上可复现;一致性判读的「可接受」标准本就依情境、无普适阈值,是方法学界三十余年来的共识(PMID 2348207 早就以「高一致率低系数」两个悖论指出单一阈值的困境;PMID 15733050 明确分级建议依赖研究目的)。最小临床重要差在不同专科算出不同数、且报告方式差异极大,已由 35 篇的系统清点证实(PMID 37933798),「最小可检测变化」与「最小重要变化」必须分开也是既有结论(PMID 16925807)。

    部分验证: 「领域效应是记账口径的产物」这一判断在仿真里成立,但我用的是合成数据、四档等距切点、两读者完全交叉。真实研究里档数、切点偏斜、读者数、稀疏轮换都会改变量级(影像方法学系统综述已指出观察者变异研究报告形式极不统一,PMID 37399082)。真实数据上倍数可能不是 1.51 与 1.12,但两个口径给出相反方向这一结构性结论不依赖具体数值——它是减法抵消的算术后果。

    待验证: ① 真实文献里把判据写成占比差形式的比例(若绝大多数写成错档率类,则「需要分领域」的普遍印象本身就是记账产物);② 三档以下量表在两种口径下的倍数差是否更小(档数越少,占比差与错档率越接近,抵消效应可能消失);③ 若两读者存在系统性偏严(仿真里加 0.4 个标准差的偏置),占比差的领域排序会反转——影像从 3.03 升到 6.38、心理量表从 3.37 降到 4.21,这意味着共同分量一旦不对称,「抵消」反而不完全,领域效应会部分回来。

    L1 网络状态:该主题尚未进入 L1 对应网络 —— CORRESPONDENCE_NETWORK.md(1351 行)与 EDGE_TABLE_science-yixue.md(63 行)逐词检索「可接受 / 门槛 / 阈值 / 判读 / 一致性 / 信度 / kappa / 最小重要差 / 领域 / 情境 / 量表 / MIC」,「可接受」仅出现 1 处在内感受量表因子结构语境;「阈值」出现于 E-061(呼吸节律判别三件套)与 P9 频散指标语境,均指生理判别阈值而非判读一致性门槛;「一致性」6 处、「信度」8 处全部落在 E-060「五脏辨证作为分组标签」的标签信度前提(标为 conflict:PMID 30383398 判 low,与 PMID 33853549 指加权系数在中低一致性大样本多类别时精度最差相冲)。边表概念最高止于生理概念,无「判读一致性门槛」这一边。故本文按硬规则封顶证据等级 C,全部结论不得读作已验证机制。

    相关阅读

    证据等级: C
    参考文献: PMID 37933798, PMID 16925807, PMID 2348207, 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。

    Is the \”Acceptable Error Margin\” a Number: The Answer Depends on Which Ruler You Use

    Evidence level: C

    Within the same domain, differences recorded via “misalignment rate” amplify by 1.51 times, whereas those recorded via “primary category proportion difference” amplify only 1.12 times—thus, the question of whether “thresholds should be domain-specific” must first address the accounting convention, not empirical data.

    Body

    I. Formulation of the Problem. The entire derivation in the previous article (#1411) rests on an untested assumption: the existence of a unified $w$, defined as “acceptable misalignment tolerance,” shared across all domains with only the value pending determination—0.05 or 0.15. A more troubling possibility is that $w$ is not a number at all, but a function varying with the interpretation task. In image interpretation, the cost of misaligning by one category is “performing an extra check,” whereas in functional assessment, the cost is “switching to a different rehabilitation protocol”; the clinical costs differ by an order of magnitude, so why use the same $w$? If $w$ truly varies by context, then the second prescription (prohibiting cross-primary-category boundary migration) must be upgraded into a stronger prohibition: prohibiting cross-domain migration, meaning even the “universal value” of 0.05 is merely a historical artifact of one domain mistakenly treated as universal.

    This suspicion is not idle speculation. In otolaryngology literature, 35 studies on minimal clinically important difference (MCID) exhibit enormous variation in calculation and reporting methods: 82.9% use anchor-based approaches, 68.6% use distribution-based approaches, only half use both, and merely 25.7% report sensitivity and specificity of that threshold (median 60.8% and 80.4%) (PMID 37933798). The same concept yields inconsistent numerical values across studies, and most studies do not even validate their accuracy. Earlier methodological reviews have settled this matter: distribution-based methods typically calculate “minimal detectable change” rather than “minimal important change,” the latter varying by anchor type, definition of minimal importance, and disease characteristics (PMID 16925807).

    II. First Simulation: Recording $w$ as Misalignment Rate. I conducted the simulation using the most intuitive convention: $w$ is defined as “misalignment rate,” i.e., the proportion of cases where two readers’ categories differ. Three “domains” were distinguished by interpretation noise (0.45, 0.75, and 1.05 latent variable standard deviations, roughly corresponding to the difficulty gaps between image interpretation, functional assessment, and psychological scales). Each domain contained 20,000 batches with 60 cases per batch.

    Result: Misalignment rate rose from 0.410 in imaging to 0.621 in psychological scales, amplifying by 1.51 times. The original judgment stated that domain-specific tables should be constructed only if the median difference in $w$ across domains exceeds 2 times; 1.51 times falls just short of this threshold. However, the 1.5-times benchmark itself was arbitrarily set; 1.51 is sufficient to demonstrate that $w$ is not a single number. Thus, the suspicion appears validated.

    III. Changing the Convention Collapses the Conclusion. Re-running the same data with $w$ redefined as the convention actually used previously—”difference in primary category proportions” (in percentage points)—yields a different outcome.

    Result: The proportion difference rises only from 3.02 percentage points in imaging to 3.37 percentage points in psychological scales, amplifying by just 1.12 times. The same domain difference and the same 1.5-times criterion are now completely unmet.

    More critically, consider the floor. The median batch-level proportion difference is identical at 5.00 percentage points across all three domains. This theoretical floor is determined by the primary category proportion $p$ and batch sample size $n$: when $p = 0.45$, the floor is 9.08 percentage points; when $p = 0.15$, it is 6.52 percentage points. The resolution of this scale depends entirely on “how full the primary category is and how large the batch is,” with almost no relation to task difficulty.

    IV. Mechanism: Why Two Scales Yield Opposite Answers. Because proportion difference involves subtraction. When two readers each deviate from the true value, harder tasks lead to larger deviations; however, the “task difficulty” component is shared—since the same batch of cases is difficult for both—so it cancels out during subtraction, leaving only the difference between the two readers. In contrast, misalignment rate is an absolute error measure; the shared component not only fails to cancel but doubles in effect.

    This explains a long-standing mystery: why earlier iterations emphasized that “the answer lies in what is not reported”—published studies almost exclusively report consistency coefficients and intervals, omitting the complete marginal distributions of each reader. Once these marginals are provided, proportion difference can be calculated; without them, one must revert to misalignment-rate-type metrics, which absorb the entire domain difficulty. Thus, “domain differences” appear large in literature largely due to accounting conventions rather than reality.

    V. Criteria Must Be Rewritten. All three original criteria fail because their conclusions depend on which quantity $w$ is recorded in. The first criterion—”construct domain-specific tables if the median difference in $w$ across domains exceeds 2 times”—does not trigger under either convention: 1.51 times fails under misalignment rate, while 1.12 times fails entirely under proportion difference—but for different reasons: one barely misses the line, the other shows no response. The second criterion—”cross-domain meta-analysis is unreliable if the proportion of cases yielding opposite conclusions across two domains exceeds 30%”—behaves differently: under proportion difference, when $w$ differs by 2 times, the contradiction rate ranges from 0.256 to 0.273 (below 30%); when differing by 3 times, it ranges from 0.475 to 0.488 (well above 30%). Under misalignment rate, regardless of whether $w$ differs by 1.5 times or 5 times, the contradiction rate stays between 0.86 and 0.98, always triggering the criterion. Thus, the same threshold represents “occasional conflict” under one convention and “permanent conflict” under another.

    Therefore, the true conclusion is neither that “$w$ must be domain-specific” nor that “$w$ can be universal,” but rather: the answer to whether “$w$ should be domain-specific” depends entirely on the accounting convention, not empirical data. You must first declare which statistical quantity $w$ is recorded in before asking whether it crosses domains. This mirrors the lesson from earlier iterations: choosing the wrong coordinate system reverses all conclusions.

    Evidence and Limitations

    Verified: The multiplicative relationships and floor formulas presented above derive from this round of simulation (20,000 batches per domain, 60 cases per batch, four-level ordered scale, latent variable model), and are arithmetically reproducible. The “acceptable” standard for consistency interpretation inherently varies by context, with no universal threshold—a consensus established over three decades in the methodology field (PMID 2348207 already pointed out the dilemma of a single threshold via two paradoxes of “high consistency, low coefficient”; PMID 15733050 explicitly recommended tiered approaches dependent on research purpose). MCID values differ across specialties and reporting methods vary greatly, confirmed by systematic review of 35 studies (PMID 37933798); the distinction between “minimal detectable change” and “minimal important change” is also an established conclusion (PMID 16925807).

    Partially Verified: The judgment that “domain effects are products of accounting conventions” holds in simulation, though I used synthetic data with four-level equally-spaced cutpoints and complete cross-over between two readers. In real studies, the number of categories, skewed cutpoints, number of readers, and sparse rotation will alter magnitudes (image methodology systematic reviews have noted extreme heterogeneity in observer variation reporting formats, PMID 37399082). On real data, the multipliers may not be exactly 1.51 and 1.12, but the structural conclusion that two conventions yield opposite directions does not depend on specific numerical values—it is an arithmetic consequence of subtraction cancellation.

    Pending Verification: ① What proportion of real literature expresses criteria in terms of proportion difference (if most are written as misalignment-rate types, then the general impression that “domain separation is needed” itself becomes a product of accounting conventions); ② Whether scales with fewer than three categories yield smaller multiplier differences under both conventions (fewer categories mean proportion difference and misalignment rate converge, potentially eliminating the cancellation effect); ③ If two readers exhibit systematic bias toward strictness (adding 0.4 standard deviations of offset in simulation), the domain ordering under proportion difference would reverse—imaging rising from 3.03 to 6.38, psychological scales falling from 3.37 to 4.21—meaning once the shared component becomes asymmetric, “cancellation” is incomplete and domain effects partially return.

    L1 Network Status: This topic has not yet entered the L1 corresponding network—CORRESPONDENCE_NETWORK.md (line 1351) and EDGE_TABLE_science-yixue.md (63 lines)—when searching term-by-term for “acceptable / threshold / tolerance / interpretation / consistency / reliability / kappa / minimal important difference / domain / context / scale / MIC.” “Acceptable” appears only once in the context of interoceptive scale factor structure; “threshold” occurs in E-061 (respiratory rhythm triad) and P9 dispersion indicator contexts, referring to physiological thresholds rather than interpretation consistency barriers; “consistency” (6 instances) and “reliability” (8 instances) all fall under E-060 “organ differentiation as grouping labels,” relying on label reliability (marked as conflict: PMID 30383398 judges low, clashing with PMID 33853549 indicating worst precision for weighted coefficients in large multi-class samples with low-medium consistency). Edge table concepts extend only to physiological concepts, lacking any “interpretation consistency threshold” edge. Hence this article caps evidence level at C per hard rules; all conclusions must not be read as verified mechanisms.

    Evidence Level: C
    References: PMID 37933798, PMID 16925807, PMID 2348207, 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.