Evidence level: C
上一篇说「加一个自带刻度的外部锚,秩就补齐」。这在代数上无懈可击,在统计上只答了一半:锚自己带误差。但把账算清之后,最要紧的一条并不是「锚太差会拖垮校正」——锚误差在总代价里只占一到两成,真正吃掉精度的是另一件事:校准只抽了一部分人,那部分人对全体的代表性才是主项。
正文
上一篇(#1403)的结论是:R 位读者只给出 R 个观测,未知量却是 R+1 个,亏的秩恒为 1,加读者补不上;只有引入一个自带单位的物理锚(皮温的摄氏度数、握力的牛顿值),把均值钉住,秩才补齐。
这是线性代数的直接推论。但它只保证「有唯一解」,不保证「解误差小」。把锚真的接进模型之后,有两件事必须分开算,而它们相差一个数量级。
一、好消息:两人之差里,锚的误差精确抵消
设读者读数 z = x + b + ε,锚 Y = k·x + η。有锚之后,每位读者的位置估计是 b̂ = z̄ − Ȳ/k。两人之差是两个 b̂ 相减,Ȳ/k 这一项在减法里被完整抵掉。
实测(n=240、4 位读者、每人重复 2 次、读者噪声 0.5):锚误差从 0 加到 0.3、1.0、3.0,两人之差的抽样标准差始终停在 0.031 至 0.034 之间,逐样本差的绝对值上界是 5.55e-17——不是「小到可以忽略」,是浮点运算的舍入底噪。
加锚对「差」这件事既没有精度代价、也没有精度收益。 那个 0.032 完全由读者噪声和病例数决定,锚误差对此一无所知。
二、绝对值那一侧:锚误差确实是全额落账的,但它只是第二项
同一次模拟里,单个读者位置估计的标准差从无锚时的 0.0223,随锚误差 0.354 涨到 0.0325(+46%),0.6 时 0.0445(翻倍),1.0 时 0.0678(三倍)。
方差可以拆成两项相加:读者项 se²/(mn) = 0.000521,锚项 s²/n。两项相等处是 s* = k·se/√m = 0.354。低于这个值锚误差基本白给,高于这个值就变成「用一把更差的尺子标定」。
但请注意这条曲线的形状:s 从 0 涨到 1.0,锚项从 0 变成 0.00417,是读者项的 8 倍——SD 从 0.0223 到 0.0678,只涨 3 倍。误差是平方相加的,最差的那一项只能占主导,不能叠加成灾。
三、真正的主项:校准只抽了一部分人
真实做法很少给全样本都上锚,通常抽一部分人做校准。这时最要紧的不是锚有多准,而是「这 n_cal 个人能不能代表全体 n 个人」。
把两件事拆开验证(n=240、se=0.5、m=2、锚误差 s=0.5):
- 所有人共用同一批校准对象(所有读者的 b̂ 都减掉同一个 Ȳ_cal):校准 10 人时差的标准差 0.99 倍,30 人 1.00 倍,60 人 1.01 倍。完全不动。 因为差值把共同项又减掉了一次。
- 每位读者各抽各的校准对象(跨中心、跨设备时这是常态):校准 10 人时 15.2 倍,30 人 8.7 倍,60 人 5.7 倍,120 人 3.6 倍,180 人 2.5 倍。
同一个锚、同一份锚误差,「共用」与「各抽」在 10 人校准点上差 15 倍。 这不是锚准不准的问题,是校准对象的身份问题。
四、而且这个代价几乎全是抽样误差,不是锚误差
把锚换成完美(s=0)再算一遍,膨胀倍数:
- 校准 10 人:13.5 倍(s=0.5 时 15.2 倍)
- 校准 30 人:7.6 倍(8.7 倍)
- 校准 60 人:5.0 倍(5.7 倍)
- 校准 120 人:3.0 倍(3.6 倍)
即便用的是零误差的完美锚,校准 10 人也要付出 13.5 倍的精度代价。 锚误差从 0 加到 0.5,在 10 人校准点上只把代价从 13.5 推到 15.2,净增 12%;在 60 人点上是 5.00 到 5.74,净增 15%;只有在接近全样本(240 人)时,锚误差才成为主项(1.00 到 1.73,净增 73%)——而那时抽样误差已经归零,剩下的自然全是锚的账。
这是本轮最实用的一条判据:校准人数不到样本的 60%,讨论「锚要多准」是抓错了主要矛盾;该讨论的是「校准要抽多少人」。
五、解析式:膨胀的平方等于 1 加上两个代价之比
上面所有数字都能用一个式子复现:
膨胀² = 1 + [ (1/n_cal − 1/n) + s²/n_cal ] ÷ (se²/(mn))
方括号里第一项是「用 n_cal 人估计全体均值」的抽样误差,第二项是锚自身的误差,分母是基线(不标定时差值的方差)。六组参数核到小数点后两位,最大偏差 0.07。
反解:在 n=240、se=0.5、m=2 的条件下,想把膨胀控制在 2 倍以内,锚完美时需要校准 175 人(占样本 73%),锚误差 0.5 时需要 218 人(91%)。想让膨胀不超过 1.5 倍,锚误差 0.5 时无解——哪怕全校准也救不回来,因为那时 s > s*,锚本身就成了最差那一项。
换句话说:所谓「抽 10% 做校准」在文献里是常规做法,在这套设计下等于把精度砍掉 15 倍。 而系统综述早就写过这条线:内部验证子样本的抽样策略研究表明,即便抽到主样本的 40%,分层随机与极值抽样也只比纯随机好 10% 到 12%,而且这点效率增益要以更高的偏倚百分比和更低的覆盖率为代价——效率不是免费的,而「抽多少」这件事比「怎么抽」重要得多。 另有方法学研究明确指出,参考测量未必是无偏金标准,其误差与被测仪器的误差可能相关,这种相关会以不可预测的方式高估或低估校正系数。
六、共模误差:这部分谁救不了
若锚误差里有一部分是全批共享的(同一台仪器、同一批试剂、同一位资深者按曲线判断),这部分不随 n 缩小,会留一个地板。
实测地板 0.13 时,样本量 60、240、960 三档的标准差是 0.1379、0.1292、0.1297——样本量涨 16 倍,只降了 5%。
加读者同样救不了。地板 0.13 下,读者从 2 位加到 16 位,绝对位置的标准差从 0.1303 只到 0.1442(反而略升,因为每人病例数被摊薄)。而两人之差那一侧纹丝不动:0.0325、0.0462、0.0643、0.0930——它只受病例数影响,完全不受共模误差影响。
这就是加锚这件事的完整账:对「差」无影响(既不加分也不减分),对「绝对值」有影响但只占一到两成,而真正的主项是校准子样本的代表性。
证据与局限
已验证: 共锚时两人之差的抽样标准差与锚误差完全无关(锚误差 0 至 3.0,标准差恒在 0.031 至 0.034,逐样本差的绝对值上界 5.55e-17);单个读者绝对位置的方差可分解为读者项 se²/(mn) 与锚项 s²/n,两项相等处 s*=k·se/√m=0.354,七组参数实测与解析最大偏差 0.0007;独锚且各读者各抽校准子样本时,差值的膨胀平方等于 1 加「抽样项 (1/n_cal−1/n) 与锚项 s²/n_cal 之和」除以基线方差,六组参数实测与解析最大偏差 0.07;共模锚误差的地板不随样本量下降(地板 0.13 时三档样本量标准差 0.1379/0.1292/0.1297);共模误差下加读者无效(2 位到 16 位绝对位置 SD 0.1303 至 0.1442,而两人差 SD 只随病例数变化)。
部分验证: 「锚误差只占总代价一到两成」这个结论是在 n=240、se=0.5、m=2、s=0.5 这一组参数下算的,参数换成小样本或高噪声时占比会上升(解析式给出占比等于 (s²/n_cal) ÷ [(1/n_cal−1/n)+s²/n_cal],n_cal 越接近 n 占比越高);「共模误差不随样本量缩小」是按「批内共同偏移 + 个体随机项」的简化两层结构模拟的,真实共享误差可能有中心—设备—批次三层,三层各自的方差占比未测;「共用同一批校准对象」和「各抽各的」是两种极端,中间态(部分重叠)的膨胀曲线未算。
待验证: ① 三位以上读者加外部锚的完整设计中,锚误差与读者噪声若正相关(锚也由人判读),有效独立精度会按 s/√(1−ρ²) 放大,而 ρ 本身从未在判读数据里测过;② 「以资深者按曲线判读为锚」这类退化情形下,锚与读者的零点是否共享同一个任意起点,未检验;③ 5pp 当量阈值在共模锚误差存在时是否仍可执行,本篇只给了地板值,没给判据失效的临界地板。
相关阅读
证据等级: C
参考文献: PMID 25627982, PMID 33878166, PMID 8793367, PMID 22468848, PMID 30402814, PMID 37704705, PMID 30383398, PMID 19538100
L1 网络状态:CORRESPONDENCE_NETWORK.md 第 202 行「E-060 标签前提:五脏辨证(作为分组标签)→ 标签信度」原记为 conflict(未裁决)——「中医医师间诊断一致性在开放人群中『低』(PMID 30383398)」与「训练可改善(PMID 19538100)」两方并列;第 223 行 KB-317 追加两处错配(样本量单位错配、精度区间错配),仍标 conflict。该边只把一致性当整体数值看,未登记「锚误差与校准抽样误差如何分别传播进该数值」这一层;全表 grep「锚 / 标定 / 零点 / 刻度 / 可识别 / 精度 / 方差 / 校准」在 591 行中仅 2 处命中且均无关(第 37 行「五脏是中医侧的锚点」、第 353 行「强度锚点=可见肌肉抽动」)。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 35 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。
Adding scale marks to a ruler is right, but adding them incorrectly makes it worse: the anchor’s error doesn’t advance into \”difference,\” but instead presses entirely onto the u…
Evidence level: C
The previous post stated that “adding an external anchor with its own scale completes the rank.” This is algebraically flawless but statistically answers only half the question: the anchor itself carries error, and this error precisely cancels out on the “how much do the two differ” side (becoming zero), but is fully accounted for on the “how far off is the zero point” side. The trouble is that what research truly needs is precisely the side that cancels out.
Main Text
The conclusion of the previous post (#1403) is: R readers provide only R observations, while the unknown quantity is R+1; the rank deficit is constantly 1, which readers cannot make up; only by introducing a physical anchor with its own unit (e.g., Celsius degrees for skin temperature, Newtons for grip strength) to fix the mean can the rank be completed.
This is a direct consequence of linear algebra, which cannot be denied. But it only may support “a unique solution exists,” not “the solution error is small.” Once the anchor is truly integrated into the model, three things happen simultaneously, and in different directions.
1. Good news: the anchor’s error precisely cancels out in the difference between two readers
Let the reader reading be $z = x + b + \varepsilon$ and the anchor be $Y = k \cdot x + \eta$. With the anchor, each reader’s position estimate is $\hat{b} = \bar{z} – \bar{Y}/k$. The difference between two readers is obtained by subtracting their $\hat{b}$ values; the term $\bar{Y}/k$ is completely cancelled out in this subtraction.
In a real test ($n=240$, 4 readers, each repeated 2 times, reader noise 0.5): as the anchor error increases from 0 to 0.3, 1.0, and 3.0, the sampling standard deviation of the difference between two readers remains constantly at 0.032; the first decimal place stays at 0.0 unchanged, and the upper bound of the absolute difference per sample is $5.55 \times 10^{-17}$—not “so small it can be ignored,” but the rounding floor noise of floating-point computation.
This is the most important sentence in this post: adding an anchor imposes no precision cost and yields no precision gain for the “difference” itself. That 0.032 is entirely determined by reader noise and case count; the anchor error knows nothing about it.
2. Bad news: on the absolute value side, the error is fully loaded
In the same simulation, the standard deviation of a single reader’s position estimate rises from 0.0224 (when the anchor error is 0) to 0.0328 (when the anchor error is 0.354, a 46% increase), and further to 0.0680 (a doubling, when the anchor error is 1.0).
Breaking this down reveals two additive terms: the reader term (0.000521) and the anchor term ($s^2/n$). These two terms are equal at $s^* = k \cdot se / \sqrt{m}$; in this example, $s^* = 0.354$. When the anchor error is below this value, reader noise dominates, and adding an anchor is essentially wasted effort; when it exceeds this value, the anchor term dominates, meaning adding an anchor is equivalent to handing over precision to a worse ruler.
And “how far off is the zero point”—this absolute value—is originally something nobody can use: it shifts arbitrarily with population composition (the true mean $\mu$). If $\mu$ shifts by 5 units, $\hat{b}$ shifts in the opposite direction by 5 units. The significance of adding an anchor lies entirely in fixing this $\mu$; the product after fixing it is precisely the half that is least needed.
3. What really goes wrong is “each reader gets its own anchor”
The cancellation in the common-anchor case occurs because everyone shares the same $\bar{Y}$, so subtraction eats up the common term. If we change to each reader calibrating against its own anchor (which is normal when crossing centers or devices), cancellation no longer occurs, and error becomes the sum of two independent terms:
When anchor error is 0.175, the standard deviation of the difference rises from 0.0323 to 0.0362 (1.12 times), requiring 269 cases to regain the original precision; at 0.354 it rises to 0.0457 (1.41 times), requiring 479 cases; at 0.600 it becomes 0.0559 (2.00 times), requiring 960 cases; at 1.000 it becomes 0.0968 (2.99 times), requiring 2148 cases.
The inflation factor is $1 + (s/s^*)^2$, which is independent of $n$, so one can only compensate by adding samples. When anchor error equals reader noise ($s = 0.354$), the sample size must double to recover; when it is three times, nine times.
This is the most practical criterion in this round: if anchor error does not fall below reader noise, adding an anchor is a trade-off of doubling the sample size to gain an unnecessary absolute value.
4. Calibration subsamples: the amplification factor is $n/n_{cal}$
In practice, it is rare to apply an anchor to the full sample; usually a subset is drawn for calibration. In this case, the anchor term becomes $s^2/n_{cal}$, while the reader term is $se^2/(mn)$; their ratio is the amplification factor. With $n=240$, $se=0.5$, $m=2$, and anchor error 0.5:
- Calibrating 10 people: SD 0.1598, which is 5 times the uncalibrated value (0.0323); converted to 5pp equivalents, it is 1.3 times worse than not calibrating
- Calibrating 30 people: SD 0.0941, 0.7 equivalents
- Calibrating 60 people: SD 0.0685, 0.54 equivalents
Systematic reviews have long noted this line: if the calibration subsample is fewer than 30 people, the confidence interval after correction widens significantly due to uncertainty in the correction coefficient itself. Other simulation studies comparing sampling strategies for internal validation subsamples found that even drawing 40% of the main sample, stratified random sampling and extreme-value sampling improve over pure random sampling by only 10% to 12%, and this efficiency gain comes at the cost of higher bias percentages and lower coverage—efficiency is not free.
5. Adding readers cannot save common-mode anchor error
If part of the anchor error is shared across the batch (e.g., same instrument, same batch of reagents, or the same senior person judging by curve), this part does not shrink with $n$ and leaves a floor.
In the real test with a floor of 0.13, the standard deviations for sample sizes of 60, 240, and 960 are 0.1371, 0.1316, and 0.1307 respectively; even though the sample size increases 16-fold, the standard deviation drops only by 5%.
Adding readers also cannot save this. When anchor error is 1.0, increasing the number of readers from 2 to 16 reduces the standard deviation only from 0.0666 to 0.0645, exactly hitting the floor of 0.0645: the $1/\sqrt{n}$ part is averaged out by the number of readers, while the other half remains unchanged.
This also explains why regression calibration literature repeatedly emphasizes that “the reference measurement is not necessarily an unbiased gold standard”—the error of the reference measurement and that of the tested instrument may be correlated, which can unpredictably overestimate or underestimate the correction coefficient; using a flawed reference measurement can lead to the true effect being underestimated by about half and the required sample size doubling.
6. When $k$ is unknown, do not perform division; directly take the intercept
All the algorithms above assume the anchor conversion coefficient $k$ is known. If $k$ is unknown, the intuitive approach is to first regress to estimate the slope and then divide; the problem is that division also divides the anchor error by $k$—when $k=0.5$, the anchor error is doubled.
The correct approach is to take the regression intercept: regress the reading against the anchor, and the intercept itself is $b$. In the real test with $k=0.5$ and anchor error 0.6, the division route yields a standard deviation of 0.0803, while the intercept route yields 0.0539, which is 1.49 times more accurate; when $k=2.0$, the two are almost identical (0.0291 vs. 0.0287).
The worry that “if $k$ is unknown, calibration cannot be done” is unnecessary, provided one uses the intercept instead of division. This corresponds to the second step of the standard three-step regression calibration in literature: regress the gold standard against observed values, and use the regression coefficients (including the intercept) themselves as the source of the correction amount.
Evidence and Limitations
Verified: When using a common anchor, the sampling standard deviation of the difference between two readers is completely independent of the anchor error (anchor error ranging from 0 to 3.0, standard deviation constantly at 0.032; the upper bound of absolute difference per sample is $5.55 \times 10^{-17}$); the variance of a single reader’s absolute position decomposes into the sum of the reader term and the anchor term, with equality at $s^ = k \cdot se / \sqrt{m} = 0.354$, and the real data matches the analytical prediction at six points ($s=0$ to $2.0$) with a maximum deviation of 0.0011; when using individual anchors, the variance inflation factor equals $1 + (s/s^)^2$, with $s=0.175/0.354/0.600/1.000$ yielding real values of 1.12/1.41/2.00/2.99 versus predicted 1.12/1.42/1.97/3.00; the amplification factor for calibration subsamples equals $n/n_{cal}$; the floor of common-mode anchor error does not decrease with sample size (with $s=0.13$, the standard deviations for three sample sizes are 0.1371/0.1316/0.1307); adding readers is ineffective under common-mode error (increasing from 2 to 16 reduces SD only from 0.0666 to 0.0645, hitting the floor); when $k$ is unknown, the intercept route outperforms the division route by 1.49 times when $k=0.5$.
Partially verified: the 5pp equivalent (0.126 standard deviations) comes from numerical solutions in the upstream of this series; this post only uses it as a precision conversion scale and did not re-derive it; the statement “common-mode error does not shrink with sample size” is based on a simplified structure of “within-batch common shift + individual random term”; real shared error may have hierarchical structure (center–device–batch three levels), and the variance proportions of each level were not measured; the intercept route outperforms the division route only under two-parameter linear relationships; if the relationship between anchor and reading is nonlinear (saturation, upper-limit truncation), the meaning of the intercept changes.
To be verified: ① In the complete design with more than three readers and an external anchor, if anchor error and reader noise are positively correlated (since the anchor is also judged by humans), effective independent precision will be amplified by $s/\sqrt{1-\rho^2}$, but $\rho$ itself has never been measured in the judgment data; ② In degenerate cases such as “using a senior person judging by curve as the anchor,” whether the zero points of the anchor and readers share the same arbitrary origin has not been tested; ③ Whether the 5pp equivalent threshold remains executable when common-mode anchor error exists; this post only provides the floor value, not the critical floor where the criterion fails.
Related Reading
- Whose ruler is bent, this question has no mathematical answer
- Reader training always practices the ruler, but no one practices the zero point
- Same batch of data, five starting points, five answers
- Resonance consciousness field research pool
Evidence level: C
References: PMID 25627982, PMID 33878166, PMID 23553407, PMID 37704705, PMID 8082971, PMID 22468848, PMID 30402814
L1 network status: In CORRESPONDENCE_NETWORK.md, line 202, “E-060 label premise: five-zang differentiation (as grouping label) → label credibility” was originally recorded as conflict (unresolved)—”consistency among TCM physicians in an open population is ‘low’ (PMID 30383398)” and “training can improve it (PMID 19538100)” stand side by side; line 223, KB-317 adds two mismatches (sample size unit mismatch, precision interval mismatch), yet still marked conflict. This edge treats consistency as an overall numerical value without registering “how anchor error propagates into the components of this number”; grepping “anchor / calibration / zero point / scale / identifiable / precision / variance” in 591 lines hits only 2 places and both are irrelevant (line 37 “five-zang is the anchor point on the TCM side”, line 353 “intensity anchor point = visible muscle twitch”). This topic has not yet entered the corresponding L1 network, and has been registered to science-yixue-l1-gaps.md (cluster “measurement consistency”, item 35).
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.
Leave a Reply